diff --git a/.github/workflows/test.yml b/.github/workflows/test.yml index f82b6e9f..f5a51ba8 100644 --- a/.github/workflows/test.yml +++ b/.github/workflows/test.yml @@ -76,22 +76,3 @@ jobs: token: ${{ secrets.CODECOV_TOKEN }} files: ./coverage.xml fail_ci_if_error: false - - test-docker-sandbox: - needs: [lint] - runs-on: ubuntu-latest - steps: - - uses: actions/checkout@v4 - - uses: astral-sh/setup-uv@v4 - with: - enable-cache: true - - name: Set up Python - uses: actions/setup-python@v5 - with: - python-version-file: "pyproject.toml" - - name: Install dependencies - run: uv sync --all-extras - - name: Build Docker image - run: docker build -t haiku-rag-slim:test -f docker/Dockerfile.slim . - - name: Run Docker integration tests - run: uv run pytest tests/agents/rlm/test_sandbox.py -v diff --git a/CHANGELOG.md b/CHANGELOG.md index efc4aa67..3acba6af 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -1,6 +1,21 @@ # Changelog ## [Unreleased] +### Changed + +- **RLM sandbox**: Replaced Docker-based code execution with [pydantic-monty](https://github.com/pydantic/monty), a minimal secure Python interpreter written in Rust. Eliminates Docker as a runtime dependency for RLM with sub-millisecond sandbox startup +- **RLM sandbox functions**: Added `get_chunk(chunk_id)` for retrieving chunk content and metadata from search results. `get_docling_document(document_id)` now returns the full document structure as a JSON dict. All sandbox functions now require `await` +- **`RLMConfig`**: Removed `docker_image` and `docker_memory_limit` fields + +### Added + +- **RLM sandbox regex functions**: `regex_findall`, `regex_sub`, `regex_search`, `regex_split` for pattern matching without LLM calls +- **`HaikuRAG.get_chunk_by_id()`**: Public method for chunk lookup by ID + +### Removed + +- **`docker_sandbox.py`**, **`runner.py`**: Docker container plumbing replaced by `sandbox.py` + ## [0.31.1] - 2026-02-20 ### Fixed diff --git a/docs/agents/rlm.md b/docs/agents/rlm.md index f4aff8e0..c339a7c0 100644 --- a/docs/agents/rlm.md +++ b/docs/agents/rlm.md @@ -5,13 +5,13 @@ The RLM agent enables complex analytical tasks by writing and executing Python c - **Aggregation**: "How many documents mention security vulnerabilities?" - **Computation**: "What's the average revenue across all quarterly reports?" - **Multi-document analysis**: "Compare the key findings between Report A and Report B" -- **Structured data extraction**: "Extract all tables from the document and summarize them" +- **Structured data extraction**: "Extract all dollar amounts and compute totals" ## How It Works 1. The agent receives a question 2. It writes Python code to explore the knowledge base -3. Code executes in a sandboxed environment with access to haiku.rag functions +3. Code executes in a sandboxed Python interpreter with access to knowledge base functions 4. The agent iterates: run code, examine results, refine approach 5. Final answer is synthesized from the gathered data @@ -52,117 +52,37 @@ async with HaikuRAG(path_to_db) as client: ) ``` -## Available Functions +## Sandbox Capabilities -Inside the sandbox, these functions are available (no imports needed): +The agent's code runs in a sandboxed Python interpreter ([pydantic-monty](https://github.com/pydantic/monty)) with access to these knowledge base functions: -### search(query, limit=10) +| Function | Description | +|----------|-------------| +| `search(query, limit)` | Hybrid search (vector + full-text) returning matching chunks with scores | +| `list_documents(limit, offset)` | List documents in the knowledge base | +| `get_document(id_or_title)` | Get full text content of a document | +| `get_chunk(chunk_id)` | Get a chunk with metadata (headings, page numbers, labels) for citations | +| `get_docling_document(document_id)` | Get the full DoclingDocument structure as a dict (texts, tables, pictures, pages) | +| `llm(prompt)` | Call an LLM for classification, summarization, or extraction | +| `regex_findall(pattern, text)`, `regex_sub(pattern, repl, text)`, `regex_search(pattern, text)`, `regex_split(pattern, text)` | Regular expression matching via Python's `re` module | -Search the knowledge base using hybrid search (vector + full-text). +When documents are pre-loaded via the `documents` parameter, they are injected as a `documents` variable accessible in the sandbox code. -```python -results = search("climate change impacts", limit=20) -for r in results: - print(r['document_title'], r['score']) - print(r['content'][:200]) -``` +### Python Features -Returns list of dicts with keys: `chunk_id`, `content`, `document_id`, `document_title`, `document_uri`, `score`, `page_numbers`, `headings` +The interpreter supports a subset of Python: variables, arithmetic, strings, f-strings, lists, dicts, tuples, sets, loops, conditionals, comprehensions, functions, async/await, try/except, and the `json` module. -### list_documents(limit=10, offset=0) +Not supported: imports (other than `json`), class definitions, generators/yield, match statements, decorators, `with` statements. For pattern matching, the agent can use the `regex_*` functions, string methods, or the `llm()` function. -List available documents in the knowledge base. +### Security -```python -docs = list_documents(limit=100) -for doc in docs: - print(doc['id'], doc['title']) -``` +Code executes in an isolated interpreter with: -Returns list of dicts with keys: `id`, `title`, `uri`, `created_at` - -### get_document(id_or_title) - -Get the full text content of a document by ID, title, or URI. - -```python -content = get_document("Q1 Report") -if content: - print(len(content), "characters") -``` - -Returns the document content as a string, or `None` if not found. - -### get_docling_document(id_or_title) - -Get the structured DoclingDocument object for advanced analysis of tables, figures, and document structure. - -```python -doc = get_docling_document("Technical Manual") -if doc: - print(f"Tables: {len(doc.tables)}") - print(f"Pictures: {len(doc.pictures)}") - - # Extract table data - for table in doc.tables: - for cell in table.data.table_cells: - print(f"Row {cell.start_row_offset_idx}, Col {cell.start_col_offset_idx}: {cell.text}") -``` - -### llm(prompt) - -Call an LLM directly for classification, summarization, or extraction tasks. - -```python -content = get_document("Q1 Report") -sentiment = llm(f"Classify the sentiment as positive, negative, or mixed: {content}") -print(sentiment) -``` - -Use this when you have content and need LLM reasoning without RAG search. - -## Pre-loaded Documents - -When documents are pre-loaded via the `documents` parameter, they're available as a `documents` variable: - -```python -# Available when documents are pre-loaded -for doc in documents: - print(doc['title'], len(doc['content'])) -``` - -Each document dict has keys: `id`, `title`, `uri`, `content` - -## Imports - -The sandbox runs in a Docker container with full Python available. Any module installed in the container image can be imported: - -```python -import re -import json -from collections import Counter - -# Extract and count patterns -results = search("error", limit=50) -error_types = [] -for r in results: - matches = re.findall(r'Error: (\w+)', r['content']) - error_types.extend(matches) - -print(Counter(error_types).most_common(10)) -``` - -The default image (`ghcr.io/ggozad/haiku.rag-slim`) includes the Python standard library. Custom images can add additional packages like `pandas` or `numpy`. - -## Docker Sandbox - -Code executes in an isolated Docker container with: - -- **Read-only database**: The LanceDB database is mounted read-only -- **Memory limits**: Configurable memory limit (default 512MB) -- **Execution timeout**: Code times out after configurable limit (default 60s) +- **No filesystem access**: Code cannot read or write files +- **No network access**: Code cannot make HTTP requests or open sockets +- **No imports**: Only the `json` module is available +- **Execution timeout**: Configurable limit (default 60s) - **Output truncation**: Large outputs are truncated to prevent memory issues -- **Container reuse**: Within a single `rlm()` call, the container stays warm for multiple code executions ## Context Filter @@ -176,11 +96,7 @@ result = await client.rlm( ) ``` -This is useful for: - -- Scoping to specific document sets -- Enforcing access control -- Limiting context for focused analysis +This is useful for scoping to specific document sets, enforcing access control, or limiting context for focused analysis. ## Configuration @@ -193,26 +109,4 @@ rlm: name: claude-sonnet-4-20250514 code_timeout: 60.0 # Max seconds for code execution max_output_chars: 50000 # Truncate output after this many chars - docker_image: "ghcr.io/ggozad/haiku.rag-slim:latest" # Container image - docker_memory_limit: "512m" # Container memory limit -``` - -### Custom Docker Image - -To add additional Python packages, create a custom Dockerfile: - -```dockerfile -FROM ghcr.io/ggozad/haiku.rag-slim:latest -RUN pip install pandas numpy -``` - -Build and configure: - -```bash -docker build -t my-rlm-image . -``` - -```yaml -rlm: - docker_image: "my-rlm-image" ``` diff --git a/docs/skills/index.md b/docs/skills/index.md index fdf689f1..e405782e 100644 --- a/docs/skills/index.md +++ b/docs/skills/index.md @@ -7,7 +7,7 @@ haiku.rag exposes its RAG capabilities as [haiku.skills](https://github.com/ggoz | Skill | Description | |-------|-------------| | [`rag`](rag.md) | Search, retrieve, and answer questions from the knowledge base | -| [`rag-rlm`](rlm.md) | Computational analysis via code execution (requires Docker) | +| [`rag-rlm`](rlm.md) | Computational analysis via code execution | ## Discovery @@ -16,7 +16,7 @@ Skills are registered as Python entrypoints under `haiku.skills`. They are disco ```bash haiku-skills list --use-entrypoints # rag — Search, retrieve and analyze documents using RAG. -# rag-rlm — Analyze documents using code execution in a Docker sandbox. +# rag-rlm — Analyze documents using code execution in a sandboxed interpreter. ``` ## Usage diff --git a/docs/skills/rlm.md b/docs/skills/rlm.md index 45f6dc14..416f5d6d 100644 --- a/docs/skills/rlm.md +++ b/docs/skills/rlm.md @@ -1,9 +1,6 @@ # RLM Skill -The RLM (Reflexion Language Model) skill provides computational analysis via code execution. It writes and runs Python code in an isolated Docker sandbox to answer questions that require computation, aggregation, or data traversal. - -!!! warning "Requires Docker" - The `analyze` tool executes code in a Docker sandbox. Docker must be running on the host machine. This skill is not suitable for Docker-deployed applications — use the [`rag`](rag.md) skill alone in those environments. +The RLM (Recursive Language Model) skill provides computational analysis via code execution. It writes and runs Python code in a sandboxed interpreter to answer questions that require computation, aggregation, or data traversal. ## `create_skill(db_path?, config?)` diff --git a/docs/tools.md b/docs/tools.md index f129c801..378d967d 100644 --- a/docs/tools.md +++ b/docs/tools.md @@ -61,7 +61,7 @@ docs = create_document_toolset(config) ### Analysis Toolset -`create_analysis_toolset()` provides computational analysis via the RLM agent (Docker sandbox). +`create_analysis_toolset()` provides computational analysis via the RLM agent. ```python from haiku.rag.tools import create_analysis_toolset diff --git a/haiku_rag_slim/haiku/rag/agents/rlm/__init__.py b/haiku_rag_slim/haiku/rag/agents/rlm/__init__.py index d5380af3..82779408 100644 --- a/haiku_rag_slim/haiku/rag/agents/rlm/__init__.py +++ b/haiku_rag_slim/haiku/rag/agents/rlm/__init__.py @@ -1,16 +1,16 @@ from haiku.rag.agents.rlm.agent import create_rlm_agent from haiku.rag.agents.rlm.dependencies import RLMContext, RLMDeps -from haiku.rag.agents.rlm.docker_sandbox import DockerSandbox, SandboxResult from haiku.rag.agents.rlm.models import CodeExecution, RLMResult from haiku.rag.agents.rlm.prompts import RLM_SYSTEM_PROMPT +from haiku.rag.agents.rlm.sandbox import Sandbox, SandboxResult __all__ = [ "CodeExecution", - "DockerSandbox", "RLMContext", "RLMDeps", "RLMResult", "RLM_SYSTEM_PROMPT", + "Sandbox", "SandboxResult", "create_rlm_agent", ] diff --git a/haiku_rag_slim/haiku/rag/agents/rlm/agent.py b/haiku_rag_slim/haiku/rag/agents/rlm/agent.py index 1b009811..4c234832 100644 --- a/haiku_rag_slim/haiku/rag/agents/rlm/agent.py +++ b/haiku_rag_slim/haiku/rag/agents/rlm/agent.py @@ -32,11 +32,10 @@ def create_rlm_agent(config: AppConfig) -> Agent[RLMDeps, RLMResult]: @agent.tool async def execute_code(ctx: RunContext[RLMDeps], code: str) -> CodeExecution: - """Execute Python code in a Docker-sandboxed environment. + """Execute Python code in a sandboxed interpreter. The code has access to haiku.rag functions (search, list_documents, - get_document, get_docling_document, llm) and any Python standard - library module. + get_document, get_chunk, llm). Use print() to output results. diff --git a/haiku_rag_slim/haiku/rag/agents/rlm/dependencies.py b/haiku_rag_slim/haiku/rag/agents/rlm/dependencies.py index 11ccaee6..02f0fdba 100644 --- a/haiku_rag_slim/haiku/rag/agents/rlm/dependencies.py +++ b/haiku_rag_slim/haiku/rag/agents/rlm/dependencies.py @@ -4,7 +4,7 @@ from typing import TYPE_CHECKING from haiku.rag.store.models import Document if TYPE_CHECKING: - from haiku.rag.agents.rlm.docker_sandbox import DockerSandbox + from haiku.rag.agents.rlm.sandbox import Sandbox @dataclass @@ -19,5 +19,5 @@ class RLMContext: class RLMDeps: """Dependencies for RLM agent.""" - sandbox: "DockerSandbox" + sandbox: "Sandbox" context: RLMContext = field(default_factory=RLMContext) diff --git a/haiku_rag_slim/haiku/rag/agents/rlm/docker_sandbox.py b/haiku_rag_slim/haiku/rag/agents/rlm/docker_sandbox.py deleted file mode 100644 index 7d91f7ca..00000000 --- a/haiku_rag_slim/haiku/rag/agents/rlm/docker_sandbox.py +++ /dev/null @@ -1,216 +0,0 @@ -"""Docker-based sandboxed execution.""" - -import asyncio -import json -import os -import subprocess -import sys -from dataclasses import dataclass -from typing import TYPE_CHECKING - -from haiku.rag.agents.rlm.dependencies import RLMContext -from haiku.rag.config.models import RLMConfig - -if TYPE_CHECKING: - from haiku.rag.client import HaikuRAG - - -@dataclass -class SandboxResult: - """Result of executing code in the sandbox.""" - - stdout: str - stderr: str - success: bool - - -class DockerSandbox: # pragma: no cover - """Execute code in a persistent Docker container. - - Use as an async context manager to manage container lifecycle: - - async with DockerSandbox(client, config, context) as sandbox: - result = await sandbox.execute("print('hello')") - result = await sandbox.execute("print('world')") - """ - - DEFAULT_IMAGE = "ghcr.io/ggozad/haiku.rag-slim:latest" - - haiku_client: "HaikuRAG" - config: RLMConfig - context: RLMContext - image: str - _process: subprocess.Popen[bytes] | None - - def __init__( - self, - client: "HaikuRAG", - config: RLMConfig, - context: RLMContext, - image: str | None = None, - ): - self.haiku_client = client - self.config = config - self.context = context - self.image = image or self.DEFAULT_IMAGE - self._process = None - - def _build_docker_cmd(self) -> list[str]: - """Build the docker run command.""" - db_path = str(self.haiku_client.store.db_path) - - env_list = ["-e", "HAIKU_DB_PATH=/data/db.lancedb"] - if self.context.filter: - env_list.extend(["-e", f"HAIKU_FILTER={self.context.filter}"]) - - ollama_host = os.environ.get("OLLAMA_HOST", "") - ollama_base_url = os.environ.get("OLLAMA_BASE_URL", "") - - if sys.platform == "darwin": - if not ollama_host or "localhost" in ollama_host: - ollama_host = "http://host.docker.internal:11434" - if not ollama_base_url or "localhost" in ollama_base_url: - ollama_base_url = "http://host.docker.internal:11434" - - if ollama_host: - env_list.extend(["-e", f"OLLAMA_HOST={ollama_host}"]) - if ollama_base_url: - env_list.extend(["-e", f"OLLAMA_BASE_URL={ollama_base_url}"]) - - for key in [ - "ANTHROPIC_API_KEY", - "OPENAI_API_KEY", - "VOYAGE_API_KEY", - "COHERE_API_KEY", - ]: - if value := os.environ.get(key): - env_list.extend(["-e", f"{key}={value}"]) - - return [ - "docker", - "run", - "--rm", - "-i", - "-v", - f"{db_path}:/data/db.lancedb:ro", - f"--memory={self.config.docker_memory_limit}", - "--network=host", - *env_list, - self.image, - "python", - "-m", - "haiku.rag.agents.rlm.runner", - ] - - async def __aenter__(self) -> "DockerSandbox": - """Start the container.""" - loop = asyncio.get_running_loop() - await loop.run_in_executor(None, self._start_container) - return self - - async def __aexit__( - self, exc_type: object, exc_val: object, exc_tb: object - ) -> None: - """Stop the container.""" - loop = asyncio.get_running_loop() - await loop.run_in_executor(None, self._stop_container) - - def _start_container(self) -> None: - """Start the persistent container process.""" - if self._process is not None: - return - - cmd = self._build_docker_cmd() - self._process = subprocess.Popen( - cmd, - stdin=subprocess.PIPE, - stdout=subprocess.PIPE, - stderr=subprocess.PIPE, - ) - - def _stop_container(self) -> None: - """Stop the container process.""" - if self._process is None: - return - - try: - if self._process.stdin: - try: - self._process.stdin.close() - except BrokenPipeError: - pass - self._process.terminate() - self._process.wait(timeout=5) - except subprocess.TimeoutExpired: - self._process.kill() - self._process.wait() - finally: - self._process = None - - async def execute(self, code: str) -> SandboxResult: - """Execute code in the container.""" - if self._process is None: - return SandboxResult( - stdout="", - stderr="Container not started. Use 'async with' context manager.", - success=False, - ) - - loop = asyncio.get_running_loop() - return await loop.run_in_executor(None, self._execute_sync, code) - - def _execute_sync(self, code: str) -> SandboxResult: - """Send code to container and read result.""" - assert self._process is not None and self._process.stdin is not None - - try: - message = json.dumps({"code": code}) - length_line = f"{len(message)}\n".encode() - self._process.stdin.write(length_line) - self._process.stdin.write(message.encode()) - self._process.stdin.flush() - - if self._process.stdout is None: - return SandboxResult( - stdout="", stderr="No stdout from container.", success=False - ) - - length_line = self._process.stdout.readline() - if not length_line: - stderr = "" - if self._process.stderr: - stderr = self._process.stderr.read().decode() - return SandboxResult( - stdout="", - stderr=stderr or "Container closed unexpectedly.", - success=False, - ) - - length = int(length_line.strip()) - response = self._process.stdout.read(length).decode() - result_data = json.loads(response) - - return SandboxResult( - stdout=result_data.get("stdout", ""), - stderr=result_data.get("stderr", ""), - success=result_data.get("success", False), - ) - - except subprocess.TimeoutExpired: - return SandboxResult( - stdout="", - stderr=f"Execution timed out after {self.config.code_timeout} seconds", - success=False, - ) - except json.JSONDecodeError as e: - return SandboxResult( - stdout="", - stderr=f"Invalid response from container: {e}", - success=False, - ) - except Exception as e: - return SandboxResult( - stdout="", - stderr=f"Execution error: {e}", - success=False, - ) diff --git a/haiku_rag_slim/haiku/rag/agents/rlm/prompts.py b/haiku_rag_slim/haiku/rag/agents/rlm/prompts.py index 10991517..cc980ddc 100644 --- a/haiku_rag_slim/haiku/rag/agents/rlm/prompts.py +++ b/haiku_rag_slim/haiku/rag/agents/rlm/prompts.py @@ -1,33 +1,52 @@ RLM_SYSTEM_PROMPT = """You are a Recursive Language Model (RLM) agent that solves complex research questions by writing and executing Python code. -IMPORTANT: You MUST use the `execute_code` tool to run Python code. The functions described below are ONLY available inside the execute_code tool - you cannot access them any other way. Always execute code to answer questions; do not just describe what code would do. +You MUST use the `execute_code` tool to run Python code. The functions described below are ONLY available inside execute_code. Always execute code to answer questions; do not just describe what code would do. -CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just use them directly: -- search("query") ✓ CORRECT -- from haiku.rag import search ✗ WRONG - will fail - -You have access to a sandboxed Python environment with these haiku.rag functions (use them directly, no imports needed): +Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just call them with `await`: +- results = await search("query") ✓ CORRECT +- import search ✗ WRONG - will fail +- results = search("query") ✗ WRONG - must use await ## Available Functions -### search(query, limit=10) -> list[dict] +### await search(query, limit=10) -> list[dict] Search the knowledge base using hybrid search (vector + full-text). Returns list of dicts with keys: chunk_id, content, document_id, document_title, document_uri, score, page_numbers, headings -### list_documents(limit=10, offset=0) -> list[dict] +### await list_documents(limit=10, offset=0) -> list[dict] List available documents in the knowledge base. Returns list of dicts with keys: id, title, uri, created_at -### get_document(id_or_title) -> str | None +### await get_document(id_or_title) -> str | None Get the full text content of a document by ID, title, or URI. Returns the document content as a string, or None if not found. -### get_docling_document(id_or_title) -> DoclingDocument | None -Get the structured DoclingDocument object for advanced analysis. -Returns a DoclingDocument object, or None if not found. -See "DoclingDocument API" section below for how to use it. +### await get_chunk(chunk_id) -> dict | None +Get a specific chunk by its ID (from search results). +Returns dict with keys: chunk_id, content, document_id, document_title, headings, page_numbers, labels +Use this to retrieve full chunk details and metadata for citation. -### llm(prompt) -> str +### await get_docling_document(document_id) -> dict | None +Get the full document structure as a dict (DoclingDocument format). +Use `list_documents()` or search results to get document IDs first. +- `texts`: list of text items, each with `text`, `label` (e.g. "title", "text", "section_header", "list_item"), and `prov` (provenance with page/bounding box) +- `tables`: list of tables, each with `data` containing `grid` (list of rows, each row a list of cells with `text`), `num_rows`, `num_cols` +- `pictures`: list of figures/images with metadata +- `pages`: page dimensions and metadata + +### await regex_findall(pattern, text) -> list[str] +Find all non-overlapping matches of a regular expression pattern in text. + +### await regex_sub(pattern, repl, text) -> str +Replace all occurrences of a regular expression pattern with a replacement string. + +### await regex_search(pattern, text) -> dict | None +Search for the first match of a pattern. Returns a dict with keys: group, groups, start, end — or None if no match. + +### await regex_split(pattern, text) -> list[str] +Split text by a regular expression pattern. + +### await llm(prompt) -> str Call an LLM directly with the given prompt. Returns the response as a string. Use this for classification, summarization, extraction, or any task where you already have the content and just need LLM reasoning. @@ -40,107 +59,69 @@ If documents were pre-loaded for this session, a `documents` variable is availab for doc in documents: print(doc['title'], len(doc['content'])) ``` -Check if it exists with: `if 'documents' in dir(): ...` +Check if it exists with: `try: documents ... except NameError: ...` -## Standard Library Modules -You can import any Python standard library module. +## Available Python Features + +The interpreter supports: variables, arithmetic, strings, f-strings, lists, dicts, tuples, sets, loops, conditionals, comprehensions, functions, async/await, `map()`, `sorted()`/`.sort(key=...)`, try/except, and the `json` module. + +Not supported: imports (other than `json`), class definitions, generators/yield, match statements, decorators, `with` statements. + +For pattern matching or text extraction, use the `regex_*` functions, string methods (`str.split`, `str.find`, `str.startswith`, `in` operator), or the `llm()` function. ## Strategy Guide -1. **Explore First**: Start by listing documents or searching to understand what's available. Document names may differ from filenames (e.g., "tbmed593.pdf" might be stored as "TB MED 593" or similar). -2. **If get_document returns None**: Use `list_documents()` to see actual document titles, or `search()` to find relevant content. +1. **Explore First**: Start by listing documents or searching to understand what's available. Document `title` is often None — use `uri` or `id` to identify documents instead. +2. **If get_document returns None**: Use `await list_documents()` to see available documents (check `uri` and `id`), or `await search()` to find relevant content. 3. **Iterative Refinement**: Run code, examine results, adjust your approach based on what you find. -4. **Use print() Liberally**: The sandbox captures stdout - print intermediate results to see what you're working with. -5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and use collections. -6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use llm(). -7. **Cite Your Sources**: Track which documents/chunks informed your answer for citation. - -## DoclingDocument API - -When you call `get_docling_document(id_or_title)`, you get a DoclingDocument object for structured document analysis. - -### Properties -- `doc.texts` - List of all text items (paragraphs, headings, etc.) -- `doc.tables` - List of all tables -- `doc.pictures` - List of all pictures/figures -- `doc.name` - Document name - -### Methods -- `doc.iterate_items(with_groups=False)` - Iterate all items with hierarchy level - Returns tuples of (item, level) where level is nesting depth -- `doc.export_to_markdown()` - Export entire document as markdown string - -### Text Item Properties -- `item.text` - The text content -- `item.label` - Type: title, paragraph, section_header, list_item, etc. (lowercase enum values) -- `item.prov` - Provenance (page numbers, bounding boxes) - -### Table Access -- `table.data.num_rows`, `table.data.num_cols` - Dimensions -- `table.data.table_cells` - List of TableCell objects -- `cell.text`, `cell.start_row_offset_idx`, `cell.start_col_offset_idx` - -### Example Usage -```python -doc = get_docling_document("My Document") - -# Get all headings -headings = [t.text for t in doc.texts if "header" in str(t.label)] - -# Iterate with structure -for item, level in doc.iterate_items(): - print(" " * level + item.text[:50]) - -# Extract table data -for table in doc.tables: - for cell in table.data.table_cells: - print(f"Row {cell.start_row_offset_idx}, Col {cell.start_col_offset_idx}: {cell.text}") -``` +4. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use `await llm()`. +5. **Cite Your Sources**: Use get_chunk() to retrieve chunk metadata for citations. Track which documents/chunks informed your answer. ## Example Patterns ### Counting documents matching a condition ```python -docs = list_documents(limit=100) +docs = await list_documents(limit=100) count = 0 for doc in docs: - content = get_document(doc['id']) + content = await get_document(doc['id']) if content and 'keyword' in content.lower(): count += 1 print(f"Found in: {doc['title']}") print(f"Total: {count}") ``` -### Aggregating data across documents +### Extracting data with regex ```python -import re numbers = [] -results = search("financial data", limit=20) +results = await search("financial data", limit=20) for r in results: - matches = re.findall(r'\\$([\\d,]+)', r['content']) - for m in matches: - numbers.append(int(m.replace(',', ''))) -print(f"Average: ${sum(numbers)/len(numbers):,.2f}") + amounts = await regex_findall(r'\\$([\\d,]+)', r['content']) + for a in amounts: + numbers.append(int(a.replace(',', ''))) +if numbers: + print(f"Average: {sum(numbers) / len(numbers)}") ``` -### Using llm() for classification +### Extracting tables from a document ```python -# Get document content -content = get_document("Q1 Report") -# Use llm() to classify sentiment -sentiment = llm(f"Classify the sentiment as positive, negative, or mixed: {content}") -print(sentiment) +docs = await list_documents(limit=10) +for d in docs: + doc = await get_docling_document(d['id']) + if doc: + tables = doc.get('tables', []) + if tables: + print(f"{d['title']}: {len(tables)} table(s)") + for i, table in enumerate(tables): + grid = table.get('data', {}).get('grid', []) + for row in grid: + cells = [cell.get('text', '') for cell in row] + print(f" Table {i}: {cells}") ``` -## Workflow - -1. **ALWAYS start by using execute_code** to explore the knowledge base -2. Run multiple code blocks as needed to gather information -3. After collecting data, provide your final answer - ## Output Format -CRITICAL: Your final response MUST be valid JSON matching this exact schema: +Your final response MUST be valid JSON matching this exact schema: ```json {"answer": "Your complete answer here as a string", "program": "Your final consolidated program here as a string"} ``` @@ -150,4 +131,4 @@ CRITICAL: Your final response MUST be valid JSON matching this exact schema: Do NOT return arbitrary JSON structures. Always use the exact format: {"answer": "...", "program": "..."} -CRITICAL: You MUST call execute_code at least once before providing your answer. Never give up without trying to execute code first.""" +You MUST call execute_code at least once before providing your answer. Never give up without trying to execute code first.""" diff --git a/haiku_rag_slim/haiku/rag/agents/rlm/runner.py b/haiku_rag_slim/haiku/rag/agents/rlm/runner.py deleted file mode 100644 index fe7be056..00000000 --- a/haiku_rag_slim/haiku/rag/agents/rlm/runner.py +++ /dev/null @@ -1,190 +0,0 @@ -"""Entry point for sandboxed code execution in Docker container.""" - -import asyncio -import json -import sys -import traceback -from io import StringIO -from typing import Any - - -def build_namespace( # pragma: no cover - client: Any, config: Any, context: Any, loop: asyncio.AbstractEventLoop -) -> dict[str, Any]: - """Build execution namespace with haiku.rag functions injected.""" - - def run_async(coro: Any) -> Any: - """Run async coroutine from sync context using thread-safe scheduling.""" - future = asyncio.run_coroutine_threadsafe(coro, loop) - return future.result(timeout=config.rlm.code_timeout) - - def search(query: str, limit: int = 10) -> list[dict]: - async def _search() -> Any: - return await client.search(query, limit=limit, filter=context.filter) - - results = run_async(_search()) - return [ - { - "chunk_id": r.chunk_id, - "content": r.content, - "document_id": r.document_id, - "document_title": r.document_title, - "document_uri": r.document_uri, - "score": r.score, - "page_numbers": r.page_numbers, - "headings": r.headings, - } - for r in results - ] - - def list_documents(limit: int = 10, offset: int = 0) -> list[dict]: - async def _list() -> Any: - return await client.list_documents( - limit=limit, offset=offset, filter=context.filter - ) - - docs = run_async(_list()) - return [ - { - "id": d.id, - "title": d.title, - "uri": d.uri, - "created_at": str(d.created_at), - } - for d in docs - ] - - def get_document(id_or_title: str) -> str | None: - async def _get() -> str | None: - doc = await client.resolve_document(id_or_title) - return doc.content if doc else None - - return run_async(_get()) - - def get_docling_document(id_or_title: str) -> Any: - async def _get() -> Any: - doc = await client.resolve_document(id_or_title) - return doc.get_docling_document() if doc else None - - return run_async(_get()) - - def llm(prompt: str) -> str: - async def _llm() -> str: - from pydantic_ai import Agent - - from haiku.rag.utils import get_model - - model = get_model(config.rlm.model, config) - agent: Agent[None, str] = Agent(model, output_type=str) - result = await agent.run(prompt) - return result.output - - return run_async(_llm()) - - namespace: dict[str, Any] = { - "search": search, - "list_documents": list_documents, - "get_document": get_document, - "get_docling_document": get_docling_document, - "llm": llm, - } - - if context.documents: - namespace["documents"] = [ - {"id": d.id, "title": d.title, "uri": d.uri, "content": d.content} - for d in context.documents - ] - - return namespace - - -def execute_code( - code: str, namespace: dict[str, Any], max_output_chars: int -) -> dict[str, Any]: - """Execute code and capture output.""" - stdout_capture = StringIO() - original_stdout = sys.stdout - - try: - sys.stdout = stdout_capture - exec(code, namespace) - stdout = stdout_capture.getvalue() - if len(stdout) > max_output_chars: - stdout = stdout[:max_output_chars] + "\n... (output truncated)" - return { - "success": True, - "stdout": stdout, - "stderr": "", - } - except Exception: - return { - "success": False, - "stdout": stdout_capture.getvalue(), - "stderr": traceback.format_exc(), - } - finally: - sys.stdout = original_stdout - - -def send_response(result: dict[str, Any]) -> None: - """Send length-prefixed JSON response.""" - response = json.dumps(result) - sys.stdout.write(f"{len(response)}\n") - sys.stdout.write(response) - sys.stdout.flush() - - -async def main() -> None: # pragma: no cover - """Main entry point for container execution. - - Runs a loop reading length-prefixed JSON messages and executing code. - """ - import concurrent.futures - import os - from pathlib import Path - - from haiku.rag.agents.rlm.dependencies import RLMContext - from haiku.rag.client import HaikuRAG - from haiku.rag.config import get_config - - config = get_config() - db_path = Path(os.environ.get("HAIKU_DB_PATH", "/data/db.lancedb")) - filter_expr = os.environ.get("HAIKU_FILTER") - context = RLMContext(filter=filter_expr) - max_output_chars = config.rlm.max_output_chars - - loop = asyncio.get_running_loop() - - async with HaikuRAG(db_path, config=config, read_only=True) as client: - namespace = build_namespace(client, config, context, loop) - - with concurrent.futures.ThreadPoolExecutor(max_workers=1) as executor: - while True: - # Read length-prefixed message - length_line = sys.stdin.readline() - if not length_line: - break - - try: - length = int(length_line.strip()) - message = sys.stdin.read(length) - request = json.loads(message) - code = request.get("code", "") - - result = await loop.run_in_executor( - executor, execute_code, code, namespace, max_output_chars - ) - send_response(result) - - except (ValueError, json.JSONDecodeError) as e: - send_response( - { - "success": False, - "stdout": "", - "stderr": f"Invalid request: {e}", - } - ) - - -if __name__ == "__main__": - asyncio.run(main()) diff --git a/haiku_rag_slim/haiku/rag/agents/rlm/sandbox.py b/haiku_rag_slim/haiku/rag/agents/rlm/sandbox.py new file mode 100644 index 00000000..8f13f5b1 --- /dev/null +++ b/haiku_rag_slim/haiku/rag/agents/rlm/sandbox.py @@ -0,0 +1,231 @@ +import json +import re +from dataclasses import dataclass +from typing import TYPE_CHECKING, Any, Literal + +import pydantic_monty + +from haiku.rag.agents.rlm.dependencies import RLMContext +from haiku.rag.config.models import AppConfig +from haiku.rag.store.compression import decompress_json + +if TYPE_CHECKING: + from haiku.rag.client import HaikuRAG + + +@dataclass +class SandboxResult: + """Result of executing code in the sandbox.""" + + stdout: str + stderr: str + success: bool + + +class Sandbox: + """Execute code in a sandboxed Python interpreter. + + Uses pydantic-monty, a minimal secure Python interpreter written in Rust. + External functions (search, list_documents, etc.) are called by Monty code + using ``await`` and resolved asynchronously on the host. + + sandbox = Sandbox(client, config, context) + result = await sandbox.execute("print('hello')") + """ + + _client: "HaikuRAG" + _config: AppConfig + _context: RLMContext + + def __init__( + self, + client: "HaikuRAG", + config: AppConfig, + context: RLMContext, + ): + self._client = client + self._config = config + self._context = context + + def _build_external_functions(self) -> dict[str, Any]: + """Build async external functions for the Monty interpreter.""" + client = self._client + config = self._config + context = self._context + + async def search(query: str, limit: int = 10) -> list[dict[str, Any]]: + results = await client.search(query, limit=limit, filter=context.filter) + return [ + { + "chunk_id": r.chunk_id, + "content": r.content, + "document_id": r.document_id, + "document_title": r.document_title, + "document_uri": r.document_uri, + "score": r.score, + "page_numbers": r.page_numbers, + "headings": r.headings, + } + for r in results + ] + + async def list_documents( + limit: int = 10, offset: int = 0 + ) -> list[dict[str, Any]]: + docs = await client.list_documents( + limit=limit, offset=offset, filter=context.filter + ) + return [ + { + "id": d.id, + "title": d.title, + "uri": d.uri, + "created_at": str(d.created_at), + } + for d in docs + ] + + async def get_document(id_or_title: str) -> str | None: + doc = await client.resolve_document(id_or_title) + return doc.content if doc else None + + async def get_chunk(chunk_id: str) -> dict[str, Any] | None: + chunk = await client.get_chunk_by_id(chunk_id) + if not chunk: + return None + meta = chunk.get_chunk_metadata() + doc_title = chunk.document_title + if not doc_title and chunk.document_id: + doc = await client.get_document_by_id(chunk.document_id) + if doc: + doc_title = doc.title + return { + "chunk_id": chunk.id, + "content": chunk.content, + "document_id": chunk.document_id, + "document_title": doc_title, + "headings": meta.headings, + "page_numbers": meta.page_numbers, + "labels": meta.labels, + } + + async def get_docling_document( + document_id: str, + ) -> dict[str, Any] | None: + doc = await client.get_document_by_id(document_id) + if not doc or not doc.docling_document: + return None + json_str = decompress_json(doc.docling_document) + return json.loads(json_str) + + async def llm(prompt: str) -> str: + from pydantic_ai import Agent + + from haiku.rag.utils import get_model + + model = get_model(config.rlm.model, config) + agent: Agent[None, str] = Agent(model, output_type=str) + result = await agent.run(prompt) + return result.output + + async def regex_findall(pattern: str, text: str) -> list[str]: + return re.findall(pattern, text) + + async def regex_sub(pattern: str, repl: str, text: str) -> str: + return re.sub(pattern, repl, text) + + async def regex_search(pattern: str, text: str) -> dict[str, Any] | None: + m = re.search(pattern, text) + if m is None: + return None + return { + "group": m.group(), + "groups": list(m.groups()), + "start": m.start(), + "end": m.end(), + } + + async def regex_split(pattern: str, text: str) -> list[str]: + return re.split(pattern, text) + + return { + "search": search, + "list_documents": list_documents, + "get_document": get_document, + "get_chunk": get_chunk, + "get_docling_document": get_docling_document, + "llm": llm, + "regex_findall": regex_findall, + "regex_sub": regex_sub, + "regex_search": regex_search, + "regex_split": regex_split, + } + + async def execute(self, code: str) -> SandboxResult: + """Execute Python code in the Monty interpreter.""" + external_fns = self._build_external_functions() + + input_names: list[str] = [] + inputs: dict[str, Any] | None = None + if self._context.documents: + input_names.append("documents") + inputs = { + "documents": [ + { + "id": d.id, + "title": d.title, + "uri": d.uri, + "content": d.content, + } + for d in self._context.documents + ] + } + + try: + monty = pydantic_monty.Monty( + code, + inputs=input_names, + external_functions=list(external_fns.keys()), + ) + except ( + pydantic_monty.MontySyntaxError, + pydantic_monty.MontyRuntimeError, + ) as e: + return SandboxResult(stdout="", stderr=str(e), success=False) + + stdout_lines: list[str] = [] + + def print_callback(_stream: Literal["stdout"], text: str) -> None: + stdout_lines.append(text) + + max_chars = self._config.rlm.max_output_chars + limits: pydantic_monty.ResourceLimits = { + "max_duration_secs": self._config.rlm.code_timeout, + } + + try: + output = await pydantic_monty.run_monty_async( + monty, + inputs=inputs, + external_functions=external_fns, + limits=limits, + print_callback=print_callback, + ) + except pydantic_monty.MontyRuntimeError as e: + stdout = "".join(stdout_lines) + if len(stdout) > max_chars: + stdout = stdout[:max_chars] + "\n... (output truncated)" + return SandboxResult(stdout=stdout, stderr=str(e), success=False) + + stdout = "".join(stdout_lines) + if output is not None: + stdout_with_output = f"{stdout}{output}" if stdout else str(output) + else: + stdout_with_output = stdout + + if len(stdout_with_output) > max_chars: + stdout_with_output = ( + stdout_with_output[:max_chars] + "\n... (output truncated)" + ) + + return SandboxResult(stdout=stdout_with_output, stderr="", success=True) diff --git a/haiku_rag_slim/haiku/rag/app.py b/haiku_rag_slim/haiku/rag/app.py index 4b2c0f15..acd229d4 100644 --- a/haiku_rag_slim/haiku/rag/app.py +++ b/haiku_rag_slim/haiku/rag/app.py @@ -354,7 +354,7 @@ class HaikuRAGApp: # pragma: no cover read_only=self.read_only, before=self.before, ) as self.client: - chunk = await self.client.chunk_repository.get_by_id(chunk_id) + chunk = await self.client.get_chunk_by_id(chunk_id) if not chunk: self.console.print(f"[red]Chunk with id {chunk_id} not found.[/red]") return diff --git a/haiku_rag_slim/haiku/rag/chat/app.py b/haiku_rag_slim/haiku/rag/chat/app.py index e0cd6142..708fae41 100644 --- a/haiku_rag_slim/haiku/rag/chat/app.py +++ b/haiku_rag_slim/haiku/rag/chat/app.py @@ -344,7 +344,7 @@ class ChatApp(App): return citation = selected_widgets[0].citation - chunk = await self.client.chunk_repository.get_by_id(citation.chunk_id) + chunk = await self.client.get_chunk_by_id(citation.chunk_id) if not chunk: return diff --git a/haiku_rag_slim/haiku/rag/client.py b/haiku_rag_slim/haiku/rag/client.py index b18f30d0..9c94a3d3 100644 --- a/haiku_rag_slim/haiku/rag/client.py +++ b/haiku_rag_slim/haiku/rag/client.py @@ -732,6 +732,17 @@ class HaikuRAG: """ return await self.document_repository.get_by_id(document_id) + async def get_chunk_by_id(self, chunk_id: str) -> Chunk | None: + """Get a chunk by its ID. + + Args: + chunk_id: The unique identifier of the chunk. + + Returns: + The Chunk instance if found, None otherwise. + """ + return await self.chunk_repository.get_by_id(chunk_id) + async def get_document_by_uri(self, uri: str) -> Document | None: """Get a document by its URI. @@ -1379,9 +1390,9 @@ class HaikuRAG: RLMResult with the answer and the final consolidated program. """ from haiku.rag.agents.rlm import ( - DockerSandbox, RLMContext, RLMDeps, + Sandbox, create_rlm_agent, ) @@ -1395,21 +1406,20 @@ class HaikuRAG: loaded_docs.append(doc) context.documents = loaded_docs if loaded_docs else None - async with DockerSandbox( + sandbox = Sandbox( client=self, - config=self._config.rlm, + config=self._config, context=context, - image=self._config.rlm.docker_image, - ) as sandbox: - deps = RLMDeps( - sandbox=sandbox, - context=context, - ) + ) + deps = RLMDeps( + sandbox=sandbox, + context=context, + ) - agent = create_rlm_agent(self._config) - result = await agent.run(question, deps=deps) + agent = create_rlm_agent(self._config) + result = await agent.run(question, deps=deps) - return result.output + return result.output async def visualize_chunk(self, chunk: Chunk) -> list: """Render page images with bounding box highlights for a chunk. diff --git a/haiku_rag_slim/haiku/rag/config/models.py b/haiku_rag_slim/haiku/rag/config/models.py index 755073e1..9e78d142 100644 --- a/haiku_rag_slim/haiku/rag/config/models.py +++ b/haiku_rag_slim/haiku/rag/config/models.py @@ -104,8 +104,6 @@ class RLMConfig(BaseModel): ) code_timeout: float = 60.0 max_output_chars: int = 50_000 - docker_image: str = "ghcr.io/ggozad/haiku.rag-slim:latest" - docker_memory_limit: str = "512m" class PictureDescriptionConfig(BaseModel): diff --git a/haiku_rag_slim/haiku/rag/inspector/widgets/search_modal.py b/haiku_rag_slim/haiku/rag/inspector/widgets/search_modal.py index b6609309..07868884 100644 --- a/haiku_rag_slim/haiku/rag/inspector/widgets/search_modal.py +++ b/haiku_rag_slim/haiku/rag/inspector/widgets/search_modal.py @@ -108,9 +108,7 @@ class SearchModal(Screen): self.chunks = [] for result in self.search_results: if result.chunk_id: - chunk = await self.client.chunk_repository.get_by_id( - result.chunk_id - ) + chunk = await self.client.get_chunk_by_id(result.chunk_id) if chunk: self.chunks.append(chunk) diff --git a/haiku_rag_slim/haiku/rag/skills/rag-rlm/SKILL.md b/haiku_rag_slim/haiku/rag/skills/rag-rlm/SKILL.md index 93253102..f59a148f 100644 --- a/haiku_rag_slim/haiku/rag/skills/rag-rlm/SKILL.md +++ b/haiku_rag_slim/haiku/rag/skills/rag-rlm/SKILL.md @@ -1,7 +1,7 @@ --- name: rag-rlm description: > - Computational analysis of the knowledge base via code execution in a Docker sandbox. + Computational analysis of the knowledge base via code execution in a sandboxed Python interpreter. Use for questions requiring counting, aggregation, statistics, data traversal, comparison across documents, or any task best answered by writing Python code. Examples: "how many pages?", "compare table 3 across documents", @@ -10,4 +10,4 @@ description: > # RLM Analysis -Use the `analyze` tool for complex analytical questions. It writes and executes Python code against the knowledge base in an isolated Docker sandbox. +Use the `analyze` tool for complex analytical questions. It writes and executes Python code against the knowledge base in a sandboxed Python interpreter. diff --git a/haiku_rag_slim/haiku/rag/tools/analysis.py b/haiku_rag_slim/haiku/rag/tools/analysis.py index b0c4a804..fd9b54a5 100644 --- a/haiku_rag_slim/haiku/rag/tools/analysis.py +++ b/haiku_rag_slim/haiku/rag/tools/analysis.py @@ -3,7 +3,7 @@ from pydantic_ai import FunctionToolset, RunContext from haiku.rag.agents.rlm.agent import create_rlm_agent from haiku.rag.agents.rlm.dependencies import RLMContext, RLMDeps -from haiku.rag.agents.rlm.docker_sandbox import DockerSandbox +from haiku.rag.agents.rlm.sandbox import Sandbox from haiku.rag.config.models import AppConfig from haiku.rag.tools.context import RAGDeps from haiku.rag.tools.filters import ( @@ -62,26 +62,25 @@ def create_analysis_toolset( rlm_context = RLMContext(filter=effective_filter) - async with DockerSandbox( + sandbox = Sandbox( client=client, - config=config.rlm, + config=config, context=rlm_context, - image=config.rlm.docker_image, - ) as sandbox: - deps = RLMDeps( - sandbox=sandbox, - context=rlm_context, - ) + ) + deps = RLMDeps( + sandbox=sandbox, + context=rlm_context, + ) - rlm_agent = create_rlm_agent(config) - result = await rlm_agent.run(task, deps=deps) + rlm_agent = create_rlm_agent(config) + result = await rlm_agent.run(task, deps=deps) - program = result.output.program + program = result.output.program - return AnalysisResult( - answer=result.output.answer, - code_executed=bool(program), - ) + return AnalysisResult( + answer=result.output.answer, + code_executed=bool(program), + ) toolset: FunctionToolset[RAGDeps] = FunctionToolset() toolset.add_function(analyze, name=tool_name) diff --git a/haiku_rag_slim/pyproject.toml b/haiku_rag_slim/pyproject.toml index 8c4ec47c..4f8bb613 100644 --- a/haiku_rag_slim/pyproject.toml +++ b/haiku_rag_slim/pyproject.toml @@ -31,6 +31,7 @@ dependencies = [ "pathspec>=1.0.3", "pydantic>=2.12.5", "pydantic-ai-slim[openai,fastmcp,logfire,ag-ui]>=1.46.0", + "pydantic-monty>=0.0.7", "python-dotenv>=1.2.1", "pyyaml>=6.0.3", "rich>=14.2.0", diff --git a/tests/agents/rlm/conftest.py b/tests/agents/rlm/conftest.py index 880d4cd5..b1df0f4c 100644 --- a/tests/agents/rlm/conftest.py +++ b/tests/agents/rlm/conftest.py @@ -1,39 +1,9 @@ -import os -import subprocess -from pathlib import Path - import pytest from haiku.rag.agents.rlm.dependencies import RLMContext -from haiku.rag.agents.rlm.docker_sandbox import DockerSandbox +from haiku.rag.agents.rlm.sandbox import Sandbox from haiku.rag.client import HaikuRAG -from haiku.rag.config.models import RLMConfig - -TEST_DOCKER_IMAGE = os.environ.get("HAIKU_TEST_DOCKER_IMAGE", "haiku-rag-slim:test") - - -@pytest.fixture(scope="session") -def test_docker_image(): - """Build and return the Docker image for testing.""" - if os.environ.get("CI"): - return TEST_DOCKER_IMAGE - - project_root = Path(__file__).parent.parent.parent.parent - dockerfile = project_root / "docker" / "Dockerfile.slim" - - if not dockerfile.exists(): - pytest.skip(f"Dockerfile.slim not found at {dockerfile}") - - result = subprocess.run( - ["docker", "build", "-t", TEST_DOCKER_IMAGE, "-f", str(dockerfile), "."], - cwd=project_root, - capture_output=True, - text=True, - ) - if result.returncode != 0: - pytest.fail(f"Failed to build Docker image:\n{result.stderr}") - - return TEST_DOCKER_IMAGE +from haiku.rag.config.models import AppConfig @pytest.fixture @@ -44,11 +14,8 @@ async def empty_client(temp_db_path): @pytest.fixture -async def docker_sandbox(empty_client, test_docker_image): - """Create a Docker sandbox for testing.""" - config = RLMConfig(docker_image=test_docker_image) +async def sandbox(empty_client): + """Create a Monty sandbox for testing.""" + config = AppConfig() context = RLMContext() - async with DockerSandbox( - client=empty_client, config=config, context=context, image=test_docker_image - ) as sandbox: - yield sandbox + return Sandbox(client=empty_client, config=config, context=context) diff --git a/tests/agents/rlm/test_agent.py b/tests/agents/rlm/test_agent.py index d6698cd7..f78af4e1 100644 --- a/tests/agents/rlm/test_agent.py +++ b/tests/agents/rlm/test_agent.py @@ -47,9 +47,7 @@ class TestClientRLMIntegration: @pytest.mark.asyncio @pytest.mark.vcr() - async def test_rlm_count_documents( - self, allow_model_requests, temp_db_path, test_docker_image - ): + async def test_rlm_count_documents(self, allow_model_requests, temp_db_path): """Test RLM agent can count documents. Agent program: @@ -59,7 +57,7 @@ class TestClientRLMIntegration: from haiku.rag.client import HaikuRAG config = AppConfig() - config.rlm.docker_image = test_docker_image + async with HaikuRAG(temp_db_path, config=config, create=True) as client: await client.create_document("First document about cats.", title="Doc 1") await client.create_document("Second document about dogs.", title="Doc 2") @@ -71,9 +69,7 @@ class TestClientRLMIntegration: @pytest.mark.asyncio @pytest.mark.vcr() - async def test_rlm_aggregation( - self, allow_model_requests, temp_db_path, test_docker_image - ): + async def test_rlm_aggregation(self, allow_model_requests, temp_db_path): """Test RLM agent can perform aggregation across documents. Agent program: @@ -95,7 +91,7 @@ class TestClientRLMIntegration: from haiku.rag.client import HaikuRAG config = AppConfig() - config.rlm.docker_image = test_docker_image + async with HaikuRAG(temp_db_path, config=config, create=True) as client: await client.create_document( "Sales report Q1: Revenue was $100,000.", title="Q1 Report" @@ -115,9 +111,7 @@ class TestClientRLMIntegration: @pytest.mark.asyncio @pytest.mark.vcr() - async def test_rlm_with_filter( - self, allow_model_requests, temp_db_path, test_docker_image - ): + async def test_rlm_with_filter(self, allow_model_requests, temp_db_path): """Test RLM agent respects filter parameter. Agent program: @@ -130,7 +124,7 @@ class TestClientRLMIntegration: from haiku.rag.client import HaikuRAG config = AppConfig() - config.rlm.docker_image = test_docker_image + async with HaikuRAG(temp_db_path, config=config, create=True) as client: await client.create_document("Cat document.", title="Cats") await client.create_document("Dog document.", title="Dogs") @@ -145,42 +139,36 @@ class TestClientRLMIntegration: @pytest.mark.asyncio @pytest.mark.vcr() - async def test_rlm_docling_document_structure( - self, allow_model_requests, temp_db_path, test_docker_image - ): - """Test RLM agent can analyze document structure using DoclingDocument. + async def test_rlm_search_and_get_chunk(self, allow_model_requests, temp_db_path): + """Test RLM agent can search and use get_chunk for citations. Agent program: - docs = list_documents(limit=20) - print(docs) - - doc = get_docling_document('') - print(doc.name) - print('tables:', len(doc.tables)) - print('pictures:', len(doc.pictures)) + results = search("content", limit=5) + for r in results: + chunk = get_chunk(r['chunk_id']) + print(chunk['document_title'], chunk['chunk_id']) """ from haiku.rag.client import HaikuRAG - pdf_path = Path("tests/data/doclaynet.pdf") config = AppConfig() - config.processing.conversion_options.do_ocr = False - config.rlm.docker_image = test_docker_image async with HaikuRAG(temp_db_path, config=config, create=True) as client: - await client.create_document_from_source(pdf_path) - - result = await client.rlm( - "How many tables are in the document? " - "Also tell me how many pictures/figures it contains." + await client.create_document( + "The quick brown fox jumps over the lazy dog.", + title="Animal Facts", ) - # The doclaynet.pdf has 1 table and 1 picture - assert "1" in result.answer + result = await client.rlm( + "Search for content about animals and tell me " + "which document it came from." + ) + + assert "Animal Facts" in result.answer @pytest.mark.asyncio @pytest.mark.vcr() async def test_rlm_semantic_analysis_with_llm( - self, allow_model_requests, temp_db_path, test_docker_image + self, allow_model_requests, temp_db_path ): """Test RLM agent can use llm() for semantic analysis combined with computation. @@ -200,7 +188,7 @@ class TestClientRLMIntegration: from haiku.rag.client import HaikuRAG config = AppConfig() - config.rlm.docker_image = test_docker_image + async with HaikuRAG(temp_db_path, config=config, create=True) as client: await client.create_document( "The new product launch exceeded expectations. Sales grew 40% " @@ -232,9 +220,7 @@ class TestClientRLMIntegration: @pytest.mark.asyncio @pytest.mark.vcr() - async def test_rlm_search_and_extract( - self, allow_model_requests, temp_db_path, test_docker_image - ): + async def test_rlm_search_and_extract(self, allow_model_requests, temp_db_path): """Test RLM agent can use search() to find content and extract information. Agent program: @@ -252,7 +238,6 @@ class TestClientRLMIntegration: pdf_path = Path("tests/data/doclaynet.pdf") config = AppConfig() config.processing.conversion_options.do_ocr = False - config.rlm.docker_image = test_docker_image async with HaikuRAG(temp_db_path, config=config, create=True) as client: await client.create_document_from_source(pdf_path) @@ -293,7 +278,7 @@ class TestClientRLMIntegration: @pytest.mark.asyncio @pytest.mark.vcr() async def test_rlm_with_preloaded_documents( - self, allow_model_requests, temp_db_path, test_docker_image + self, allow_model_requests, temp_db_path ): """Test RLM agent can use pre-loaded documents variable. @@ -307,7 +292,7 @@ class TestClientRLMIntegration: from haiku.rag.client import HaikuRAG config = AppConfig() - config.rlm.docker_image = test_docker_image + async with HaikuRAG(temp_db_path, config=config, create=True) as client: await client.create_document( "The company was founded in 1985 by Jane Smith.", diff --git a/tests/agents/rlm/test_runner.py b/tests/agents/rlm/test_runner.py deleted file mode 100644 index 18104917..00000000 --- a/tests/agents/rlm/test_runner.py +++ /dev/null @@ -1,50 +0,0 @@ -import json -from io import StringIO - -from haiku.rag.agents.rlm.runner import execute_code, send_response - - -def test_execute_code_success(): - namespace: dict = {} - result = execute_code("x = 1 + 1", namespace, max_output_chars=1000) - assert result["success"] is True - assert result["stderr"] == "" - - -def test_execute_code_stdout_capture(): - namespace: dict = {} - result = execute_code("print('hello')", namespace, max_output_chars=1000) - assert result["success"] is True - assert "hello" in result["stdout"] - - -def test_execute_code_exception(): - namespace: dict = {} - result = execute_code("raise ValueError('boom')", namespace, max_output_chars=1000) - assert result["success"] is False - assert "ValueError" in result["stderr"] - assert "boom" in result["stderr"] - - -def test_execute_code_output_truncation(): - namespace: dict = {} - code = "print('x' * 100)" - result = execute_code(code, namespace, max_output_chars=10) - assert result["success"] is True - assert "truncated" in result["stdout"] - assert len(result["stdout"]) < 100 - - -def test_send_response(monkeypatch): - buf = StringIO() - monkeypatch.setattr("sys.stdout", buf) - - payload = {"success": True, "stdout": "hi", "stderr": ""} - send_response(payload) - - output = buf.getvalue() - lines = output.split("\n", 1) - length = int(lines[0]) - body = lines[1] - assert json.loads(body) == payload - assert length == len(json.dumps(payload)) diff --git a/tests/agents/rlm/test_sandbox.py b/tests/agents/rlm/test_sandbox.py index 50223b5b..8ab1c62d 100644 --- a/tests/agents/rlm/test_sandbox.py +++ b/tests/agents/rlm/test_sandbox.py @@ -1,12 +1,12 @@ -import os from pathlib import Path import pytest from haiku.rag.agents.rlm.dependencies import RLMContext -from haiku.rag.agents.rlm.docker_sandbox import DockerSandbox, SandboxResult +from haiku.rag.agents.rlm.sandbox import Sandbox, SandboxResult from haiku.rag.client import HaikuRAG -from haiku.rag.config.models import RLMConfig +from haiku.rag.config.models import AppConfig +from haiku.rag.store.models import Document @pytest.fixture(scope="module") @@ -14,103 +14,83 @@ def vcr_cassette_dir(): return str(Path(__file__).parent.parent.parent / "cassettes" / "test_sandbox") -def is_docker_available() -> bool: - """Check if Docker daemon is available.""" - try: - import subprocess +class TestSandboxBasics: + """Test basic sandbox functionality.""" - result = subprocess.run(["docker", "info"], capture_output=True, timeout=5) - return result.returncode == 0 - except Exception: - return False - - -docker_required = pytest.mark.skipif( - not is_docker_available(), - reason="Docker daemon not available", -) - - -@pytest.mark.integration -class TestDockerSandboxBasics: - """Test basic Docker sandbox functionality.""" - - @docker_required @pytest.mark.asyncio - async def test_execute_simple_code(self, docker_sandbox): + async def test_execute_simple_code(self, sandbox): """Test executing simple code in the sandbox.""" - result = await docker_sandbox.execute("print('hello world')") + result = await sandbox.execute("print('hello world')") assert isinstance(result, SandboxResult) assert result.success assert "hello world" in result.stdout assert result.stderr == "" - -@pytest.mark.integration -class TestDockerSandboxErrors: - """Test error handling in Docker sandbox.""" - - @docker_required @pytest.mark.asyncio - async def test_syntax_error(self, docker_sandbox): + async def test_execute_expression_output(self, sandbox): + """Test that expression values are captured.""" + result = await sandbox.execute("1 + 2") + assert result.success + assert "3" in result.stdout + + @pytest.mark.asyncio + async def test_execute_print_and_expression(self, sandbox): + """Test print output combined with expression value.""" + result = await sandbox.execute("print('hello')\n42") + assert result.success + assert "hello" in result.stdout + assert "42" in result.stdout + + +class TestSandboxErrors: + """Test error handling in sandbox.""" + + @pytest.mark.asyncio + async def test_syntax_error(self, sandbox): """Test that syntax errors are reported.""" - result = await docker_sandbox.execute("def foo(") + result = await sandbox.execute("def foo(") assert not result.success - assert "SyntaxError" in result.stderr + assert result.stderr != "" - @docker_required @pytest.mark.asyncio - async def test_runtime_error(self, docker_sandbox): + async def test_runtime_error(self, sandbox): """Test that runtime errors are reported.""" - result = await docker_sandbox.execute("x = 1/0") + result = await sandbox.execute("x = 1/0") assert not result.success assert "ZeroDivisionError" in result.stderr - @docker_required @pytest.mark.asyncio - async def test_name_error(self, docker_sandbox): + async def test_name_error(self, sandbox): """Test that name errors are reported.""" - result = await docker_sandbox.execute("print(undefined_variable)") + result = await sandbox.execute("print(undefined_variable)") assert not result.success assert "NameError" in result.stderr - @docker_required @pytest.mark.asyncio - async def test_missing_image(self, temp_db_path): - """Test error when Docker image is not found.""" - async with HaikuRAG(temp_db_path, create=True) as client: - config = RLMConfig(docker_image="nonexistent-image:v999.999.999") - context = RLMContext() - async with DockerSandbox( - client=client, config=config, context=context, image=config.docker_image - ) as sandbox: - result = await sandbox.execute("print('hello')") - assert not result.success - assert ( - "not found" in result.stderr.lower() - or "error" in result.stderr.lower() - ) + async def test_multi_module_import(self, sandbox): + """Test that unsupported multi-module imports are caught gracefully.""" + result = await sandbox.execute("import json, string") + assert not result.success + assert result.stderr != "" -@pytest.mark.integration -class TestDockerSandboxHaikuRAG: - """Test haiku.rag functions in Docker sandbox.""" +class TestSandboxHaikuRAG: + """Test haiku.rag functions in sandbox.""" - @docker_required @pytest.mark.asyncio - async def test_list_documents_empty(self, docker_sandbox): + async def test_list_documents_empty(self, sandbox): """Test list_documents returns empty list for empty database.""" - result = await docker_sandbox.execute( - "docs = list_documents()\nprint(type(docs).__name__, len(docs))" + result = await sandbox.execute( + "docs = await list_documents()\nprint(type(docs).__name__, len(docs))" ) assert result.success assert "list 0" in result.stdout - @docker_required @pytest.mark.asyncio @pytest.mark.vcr() - async def test_list_documents_with_data(self, temp_db_path, test_docker_image): + async def test_list_documents_with_data(self, temp_db_path): """Test list_documents returns documents when populated.""" + config = AppConfig() async with HaikuRAG(temp_db_path, create=True) as client: await client.create_document( content="Test content", @@ -118,27 +98,22 @@ class TestDockerSandboxHaikuRAG: title="Test Document", ) - config = RLMConfig(docker_image=test_docker_image) context = RLMContext() - async with DockerSandbox( - client=client, config=config, context=context, image=test_docker_image - ) as sandbox: - result = await sandbox.execute( - "docs = list_documents()\nprint(len(docs))\nprint(docs[0]['title'])" - ) - assert result.success - assert "1" in result.stdout - assert "Test Document" in result.stdout + sb = Sandbox(client=client, config=config, context=context) + result = await sb.execute( + "docs = await list_documents()\n" + "print(len(docs))\n" + "print(docs[0]['title'])" + ) + assert result.success + assert "1" in result.stdout + assert "Test Document" in result.stdout - @docker_required @pytest.mark.asyncio @pytest.mark.vcr() - @pytest.mark.skipif( - os.environ.get("CI") == "true", - reason="Requires Ollama - VCR can't capture calls from inside Docker", - ) - async def test_search_with_data(self, temp_db_path, test_docker_image): + async def test_search_with_data(self, temp_db_path): """Test search function works.""" + config = AppConfig() async with HaikuRAG(temp_db_path, create=True) as client: await client.create_document( content="The quick brown fox jumps over the lazy dog.", @@ -146,26 +121,22 @@ class TestDockerSandboxHaikuRAG: title="Animals", ) - config = RLMConfig(docker_image=test_docker_image) context = RLMContext() - async with DockerSandbox( - client=client, config=config, context=context, image=test_docker_image - ) as sandbox: - result = await sandbox.execute( - "results = search('fox', limit=5)\n" - "print(len(results))\n" - "if results:\n" - " print('fox' in results[0]['content'].lower())" - ) - assert result.success - # Search should return at least one result - assert "True" in result.stdout or "1" in result.stdout + sb = Sandbox(client=client, config=config, context=context) + result = await sb.execute( + "results = await search('fox', limit=5)\n" + "print(len(results))\n" + "if results:\n" + " print('fox' in results[0]['content'].lower())" + ) + assert result.success + assert "True" in result.stdout or "1" in result.stdout - @docker_required @pytest.mark.asyncio @pytest.mark.vcr() - async def test_get_document(self, temp_db_path, test_docker_image): + async def test_get_document(self, temp_db_path): """Test get_document function.""" + config = AppConfig() async with HaikuRAG(temp_db_path, create=True) as client: doc = await client.create_document( content="Content about foxes and dogs.", @@ -173,40 +144,149 @@ class TestDockerSandboxHaikuRAG: title="Fox Document", ) - config = RLMConfig(docker_image=test_docker_image) context = RLMContext() - async with DockerSandbox( - client=client, config=config, context=context, image=test_docker_image - ) as sandbox: - result = await sandbox.execute( - f"content = get_document('{doc.id}')\n" - "print('foxes' in content.lower() if content else 'None')" - ) - assert result.success - assert "True" in result.stdout + sb = Sandbox(client=client, config=config, context=context) + result = await sb.execute( + f"content = await get_document('{doc.id}')\n" + "print('foxes' in content.lower() if content else 'None')" + ) + assert result.success + assert "True" in result.stdout - @docker_required @pytest.mark.asyncio - async def test_get_document_not_found(self, docker_sandbox): + async def test_get_document_not_found(self, sandbox): """Test get_document returns None for missing document.""" - result = await docker_sandbox.execute( - "content = get_document('nonexistent-id')\nprint(content is None)" + result = await sandbox.execute( + "content = await get_document('nonexistent-id')\nprint(content is None)" + ) + assert result.success + assert "True" in result.stdout + + @pytest.mark.asyncio + @pytest.mark.vcr() + async def test_get_chunk(self, temp_db_path): + """Test get_chunk function returns chunk with metadata.""" + config = AppConfig() + async with HaikuRAG(temp_db_path, create=True) as client: + await client.create_document( + content="Content about foxes and dogs.", + uri="test://doc", + title="Fox Document", + ) + + context = RLMContext() + sb = Sandbox(client=client, config=config, context=context) + # First search to get a chunk_id + result = await sb.execute( + "results = await search('foxes', limit=1)\n" + "chunk_id = results[0]['chunk_id']\n" + "chunk = await get_chunk(chunk_id)\n" + "print(chunk['document_title'])\n" + "print('content' in chunk)" + ) + assert result.success + assert "Fox Document" in result.stdout + assert "True" in result.stdout + + @pytest.mark.asyncio + async def test_get_chunk_not_found(self, sandbox): + """Test get_chunk returns None for missing chunk.""" + result = await sandbox.execute( + "chunk = await get_chunk('nonexistent-id')\nprint(chunk is None)" ) assert result.success assert "True" in result.stdout -@pytest.mark.integration -class TestDockerSandboxContextFilter: +class TestSandboxExternalFunctionEdgeCases: + """Test edge cases in external function dispatch.""" + + @pytest.mark.asyncio + async def test_unknown_external_function(self, sandbox): + """Test that calling an unregistered external function resumes with KeyError.""" + original_build = sandbox._build_external_functions + + def patched_build(): + fns = original_build() + fns["search"] = None + return fns + + sandbox._build_external_functions = patched_build + + result = await sandbox.execute( + "try:\n" + " await search('hello')\n" + "except:\n" + " print('caught')\n" + "print('done')" + ) + assert result.success + assert "caught" in result.stdout + assert "done" in result.stdout + + @pytest.mark.asyncio + async def test_external_function_raises_exception(self, sandbox): + """Test that exceptions from async external functions surface as errors. + + With run_monty_async, exceptions from async external functions + propagate as MontyRuntimeError rather than being catchable inside + Monty's try/except. + """ + original_build = sandbox._build_external_functions + + def patched_build(): + fns = original_build() + + async def failing_search(*args, **kwargs): + raise ValueError("external error") + + fns["search"] = failing_search + return fns + + sandbox._build_external_functions = patched_build + + result = await sandbox.execute("await search('hello')") + assert not result.success + assert "external error" in result.stderr + + +class TestSandboxOutputTruncation: + """Test output truncation behavior.""" + + @pytest.mark.asyncio + async def test_truncate_stdout_on_runtime_error(self, empty_client): + """Test stdout is truncated when a runtime error occurs after large output.""" + config = AppConfig() + config.rlm.max_output_chars = 20 + context = RLMContext() + sb = Sandbox(client=empty_client, config=config, context=context) + result = await sb.execute("print('a' * 100)\nx = 1/0") + assert not result.success + assert "ZeroDivisionError" in result.stderr + assert result.stdout.endswith("... (output truncated)") + assert len(result.stdout) < 100 + + @pytest.mark.asyncio + async def test_truncate_successful_output(self, empty_client): + """Test output is truncated on successful execution with large output.""" + config = AppConfig() + config.rlm.max_output_chars = 20 + context = RLMContext() + sb = Sandbox(client=empty_client, config=config, context=context) + result = await sb.execute("print('b' * 100)") + assert result.success + assert result.stdout.endswith("... (output truncated)") + assert len(result.stdout) < 100 + + +class TestSandboxContextFilter: """Test context filter is applied.""" - @docker_required @pytest.mark.asyncio @pytest.mark.vcr() - async def test_filter_applied_to_list_documents( - self, temp_db_path, test_docker_image - ): + async def test_filter_applied_to_list_documents(self, temp_db_path): """Test that context filter is passed to list_documents.""" + config = AppConfig() async with HaikuRAG(temp_db_path, create=True) as client: await client.create_document( content="Public content", @@ -219,33 +299,165 @@ class TestDockerSandboxContextFilter: title="Private Doc", ) - config = RLMConfig(docker_image=test_docker_image) context = RLMContext(filter="uri LIKE 'public://%'") - async with DockerSandbox( - client=client, config=config, context=context, image=test_docker_image - ) as sandbox: - result = await sandbox.execute( - "docs = list_documents()\n" - "print(len(docs))\n" - "if docs:\n" - " print(docs[0]['title'])" - ) - assert result.success - assert "1" in result.stdout - assert "Public Doc" in result.stdout - assert "Private Doc" not in result.stdout + sb = Sandbox(client=client, config=config, context=context) + result = await sb.execute( + "docs = await list_documents()\n" + "print(len(docs))\n" + "if docs:\n" + " print(docs[0]['title'])" + ) + assert result.success + assert "1" in result.stdout + assert "Public Doc" in result.stdout + assert "Private Doc" not in result.stdout -@pytest.mark.integration -class TestDockerSandboxPreloadedDocuments: +class TestSandboxPreloadedDocuments: """Test pre-loaded documents context variable.""" - @docker_required @pytest.mark.asyncio - async def test_documents_variable_not_available_without_preload( - self, docker_sandbox - ): + async def test_documents_variable_not_available_without_preload(self, sandbox): """documents variable is not available when context.documents is None.""" - result = await docker_sandbox.execute("print(documents)") + result = await sandbox.execute("print(documents)") assert not result.success assert "NameError" in result.stderr + + @pytest.mark.asyncio + async def test_documents_variable_available_with_preload(self, empty_client): + """documents variable is available when context.documents is set.""" + config = AppConfig() + docs = [ + Document(id="1", content="Content A", title="Doc A", uri="a://1"), + Document(id="2", content="Content B", title="Doc B", uri="b://2"), + ] + context = RLMContext(documents=docs) + sb = Sandbox(client=empty_client, config=config, context=context) + result = await sb.execute( + "print(len(documents))\n" + "print(documents[0]['title'])\n" + "print(documents[1]['title'])" + ) + assert result.success + assert "2" in result.stdout + assert "Doc A" in result.stdout + assert "Doc B" in result.stdout + + +class TestSandboxDoclingDocument: + """Test get_docling_document() external function.""" + + @pytest.mark.asyncio + async def test_returns_none_for_missing_document(self, sandbox): + """get_docling_document returns None for a non-existent document.""" + result = await sandbox.execute( + "doc = await get_docling_document('nonexistent-id')\nprint(doc is None)" + ) + assert result.success + assert "True" in result.stdout + + @pytest.mark.asyncio + @pytest.mark.vcr() + async def test_returns_dict_for_document_with_docling_data(self, temp_db_path): + """get_docling_document returns a dict for a document with docling data.""" + config = AppConfig() + async with HaikuRAG(temp_db_path, create=True) as client: + doc = await client.create_document( + content="Docling processed content", + uri="test://docling", + title="Docling Doc", + ) + + context = RLMContext() + sb = Sandbox(client=client, config=config, context=context) + result = await sb.execute( + f"doc = await get_docling_document('{doc.id}')\n" + "print(type(doc).__name__)\n" + "print(doc['name'])\n" + "print('texts' in doc)" + ) + assert result.success + assert "dict" in result.stdout + assert "True" in result.stdout + + +class TestSandboxRegex: + """Test regex external functions.""" + + @pytest.mark.asyncio + async def test_regex_findall(self, sandbox): + """regex_findall extracts all matches.""" + result = await sandbox.execute( + r"matches = await regex_findall(r'\d+', 'abc 123 def 456')" + "\nprint(matches)" + ) + assert result.success + assert "['123', '456']" in result.stdout + + @pytest.mark.asyncio + async def test_regex_sub(self, sandbox): + """regex_sub replaces matches.""" + result = await sandbox.execute( + r"out = await regex_sub(r'\d+', 'X', 'abc 123 def 456')" + "\nprint(out)" + ) + assert result.success + assert "abc X def X" in result.stdout + + @pytest.mark.asyncio + async def test_regex_search_match(self, sandbox): + """regex_search returns match dict when pattern matches.""" + result = await sandbox.execute( + r"m = await regex_search(r'(\d+)', 'abc 123')" + "\nprint(m['group'])" + "\nprint(m['start'])" + "\nprint(m['end'])" + ) + assert result.success + assert "123" in result.stdout + assert "4" in result.stdout + assert "7" in result.stdout + + @pytest.mark.asyncio + async def test_regex_search_no_match(self, sandbox): + """regex_search returns None when pattern doesn't match.""" + result = await sandbox.execute( + r"m = await regex_search(r'\d+', 'abc')" + "\nprint(m is None)" + ) + assert result.success + assert "True" in result.stdout + + @pytest.mark.asyncio + async def test_regex_split(self, sandbox): + """regex_split splits on pattern.""" + result = await sandbox.execute( + "out = await regex_split(',', 'a,b,,c')\nprint(out)" + ) + assert result.success + assert "['a', 'b', '', 'c']" in result.stdout + + @pytest.mark.asyncio + async def test_regex_invalid_pattern(self, sandbox): + """Invalid regex pattern surfaces as an error.""" + result = await sandbox.execute("await regex_findall('[invalid', 'text')") + assert not result.success + assert result.stderr != "" + + +class TestSandboxLLM: + """Test llm() external function.""" + + @pytest.mark.asyncio + @pytest.mark.vcr() + async def test_llm_function(self, allow_model_requests, empty_client): + """Test llm() calls the model and returns a string.""" + config = AppConfig() + context = RLMContext() + sb = Sandbox(client=empty_client, config=config, context=context) + result = await sb.execute( + "answer = await llm('What is 2 + 2? Reply with just the number.')\n" + "print(answer)" + ) + assert result.success + assert "4" in result.stdout diff --git a/tests/cassettes/test_rlm/TestClientRLMIntegration.test_rlm_aggregation.yaml b/tests/cassettes/test_rlm/TestClientRLMIntegration.test_rlm_aggregation.yaml index cade833f..f2b1b2ce 100644 --- a/tests/cassettes/test_rlm/TestClientRLMIntegration.test_rlm_aggregation.yaml +++ b/tests/cassettes/test_rlm/TestClientRLMIntegration.test_rlm_aggregation.yaml @@ -128,7 +128,7 @@ interactions: connection: - keep-alive content-length: - - '7790' + - '7341' content-type: - application/json host: @@ -141,32 +141,33 @@ interactions: IMPORTANT: You MUST use the `execute_code` tool to run Python code. The functions described below are ONLY available inside the execute_code tool - you cannot access them any other way. Always execute code to answer questions; do not just describe what code would do. - CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just use them directly: - - search("query") ✓ CORRECT + CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just call them with `await`: + - results = await search("query") ✓ CORRECT - from haiku.rag import search ✗ WRONG - will fail + - results = search("query") ✗ WRONG - must use await - You have access to a sandboxed Python environment with these haiku.rag functions (use them directly, no imports needed): + You have access to a sandboxed Python interpreter with these haiku.rag functions (use them directly with `await`, no imports needed): ## Available Functions - ### search(query, limit=10) -> list[dict] + ### await search(query, limit=10) -> list[dict] Search the knowledge base using hybrid search (vector + full-text). Returns list of dicts with keys: chunk_id, content, document_id, document_title, document_uri, score, page_numbers, headings - ### list_documents(limit=10, offset=0) -> list[dict] + ### await list_documents(limit=10, offset=0) -> list[dict] List available documents in the knowledge base. Returns list of dicts with keys: id, title, uri, created_at - ### get_document(id_or_title) -> str | None + ### await get_document(id_or_title) -> str | None Get the full text content of a document by ID, title, or URI. Returns the document content as a string, or None if not found. - ### get_docling_document(id_or_title) -> DoclingDocument | None - Get the structured DoclingDocument object for advanced analysis. - Returns a DoclingDocument object, or None if not found. - See "DoclingDocument API" section below for how to use it. + ### await get_chunk(chunk_id) -> dict | None + Get a specific chunk by its ID (from search results). + Returns dict with keys: chunk_id, content, document_id, document_title, headings, page_numbers, labels + Use this to retrieve full chunk details and metadata for citation. - ### llm(prompt) -> str + ### await llm(prompt) -> str Call an LLM directly with the given prompt. Returns the response as a string. Use this for classification, summarization, extraction, or any task where you already have the content and just need LLM reasoning. @@ -181,93 +182,64 @@ interactions: ``` Check if it exists with: `if 'documents' in dir(): ...` - ## Standard Library Modules - You can import any Python standard library module. + ## Available Python Features + + The interpreter supports: variables, arithmetic, strings, f-strings, lists, dicts, tuples, sets, loops, conditionals, comprehensions, functions, async/await, `map()`, `sorted()`/`.sort(key=...)`, try/except, and the `json` module. + + Not supported: imports (other than `json`), class definitions, generators/yield, match statements, decorators, `with` statements. + + For pattern matching or text extraction, use string methods (`str.split`, `str.find`, `str.startswith`, `in` operator) or the `llm()` function. ## Strategy Guide 1. **Explore First**: Start by listing documents or searching to understand what's available. Document names may differ from filenames (e.g., "tbmed593.pdf" might be stored as "TB MED 593" or similar). - 2. **If get_document returns None**: Use `list_documents()` to see actual document titles, or `search()` to find relevant content. + 2. **If get_document returns None**: Use `await list_documents()` to see actual document titles, or `await search()` to find relevant content. 3. **Iterative Refinement**: Run code, examine results, adjust your approach based on what you find. - 4. **Use print() Liberally**: The REPL captures stdout - print intermediate results to see what you're working with. - 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and use collections. - 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use llm(). - 7. **Cite Your Sources**: Track which documents/chunks informed your answer for citation. - - ## DoclingDocument API - - When you call `get_docling_document(id_or_title)`, you get a DoclingDocument object for structured document analysis. - - ### Properties - - `doc.texts` - List of all text items (paragraphs, headings, etc.) - - `doc.tables` - List of all tables - - `doc.pictures` - List of all pictures/figures - - `doc.name` - Document name - - ### Methods - - `doc.iterate_items(with_groups=False)` - Iterate all items with hierarchy level - Returns tuples of (item, level) where level is nesting depth - - `doc.export_to_markdown()` - Export entire document as markdown string - - ### Text Item Properties - - `item.text` - The text content - - `item.label` - Type: title, paragraph, section_header, list_item, etc. (lowercase enum values) - - `item.prov` - Provenance (page numbers, bounding boxes) - - ### Table Access - - `table.data.num_rows`, `table.data.num_cols` - Dimensions - - `table.data.table_cells` - List of TableCell objects - - `cell.text`, `cell.start_row_offset_idx`, `cell.start_col_offset_idx` - - ### Example Usage - ```python - doc = get_docling_document("My Document") - - # Get all headings - headings = [t.text for t in doc.texts if "header" in str(t.label)] - - # Iterate with structure - for item, level in doc.iterate_items(): - print(" " * level + item.text[:50]) - - # Extract table data - for table in doc.tables: - for cell in table.data.table_cells: - print(f"Row {cell.start_row_offset_idx}, Col {cell.start_col_offset_idx}: {cell.text}") - ``` + 4. **Use print() Liberally**: The sandbox captures stdout - print intermediate results to see what you're working with. + 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and data structures. + 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use `await llm()`. + 7. **Cite Your Sources**: Use get_chunk() to retrieve chunk metadata for citations. Track which documents/chunks informed your answer. ## Example Patterns ### Counting documents matching a condition ```python - docs = list_documents(limit=100) + docs = await list_documents(limit=100) count = 0 for doc in docs: - content = get_document(doc['id']) + content = await get_document(doc['id']) if content and 'keyword' in content.lower(): count += 1 print(f"Found in: {doc['title']}") print(f"Total: {count}") ``` - ### Aggregating data across documents + ### Extracting data with llm() ```python - import re numbers = [] - results = search("financial data", limit=20) + results = await search("financial data", limit=20) for r in results: - matches = re.findall(r'\$([\d,]+)', r['content']) - for m in matches: - numbers.append(int(m.replace(',', ''))) - print(f"Average: ${sum(numbers)/len(numbers):,.2f}") + extracted = await llm(f"Extract all dollar amounts from this text as a comma-separated list of numbers (no $ signs): {r['content']}") + for part in extracted.split(','): + part = part.strip().replace(',', '') + if part.isdigit(): + numbers.append(int(part)) + if numbers: + print(f"Average: {sum(numbers) / len(numbers)}") + ``` + + ### Using search results with get_chunk for citations + ```python + results = await search("safety requirements", limit=5) + for r in results: + chunk = await get_chunk(r['chunk_id']) + print(f"From '{chunk['document_title']}', page {chunk['page_numbers']}: {chunk['content'][:100]}") ``` ### Using llm() for classification ```python - # Get document content - content = get_document("Q1 Report") - # Use llm() to classify sentiment - sentiment = llm(f"Classify the sentiment as positive, negative, or mixed: {content}") + content = await get_document("Q1 Report") + sentiment = await llm(f"Classify the sentiment as positive, negative, or mixed: {content}") print(sentiment) ``` @@ -300,11 +272,10 @@ interactions: tools: - function: description: |- - Execute Python code in a Docker-sandboxed environment. + Execute Python code in a sandboxed interpreter. The code has access to haiku.rag functions (search, list_documents, - get_document, get_docling_document, llm) and any Python standard - library module. + get_document, get_chunk, llm). Use print() to output results. @@ -344,7 +315,7 @@ interactions: response: headers: content-length: - - '777' + - '809' content-type: - application/json parsed_body: @@ -353,26 +324,27 @@ interactions: index: 0 message: content: '' - reasoning: Need to find quarterly reports, extract revenue amounts, sum. Likely documents titled like "Q1 Report", - "Q2 Report"... Let's search "quarterly report revenue". + reasoning: We need to find quarterly reports documents and extract revenue figures. Likely documents include Q1, + Q2, etc. Search for "quarterly report revenue". role: assistant tool_calls: - function: - arguments: '{"code":"results = search(\"quarterly report revenue\", limit=20)\nprint(len(results))\nfor r in - results[:5]:\n print(r[''document_title''], r[''page_numbers''], r[''score''])\n"}' + arguments: '{"code":"# Search for quarterly report revenue\nresults = await search(\"quarterly report revenue\", + limit=20)\nprint(len(results))\nfor r in results[:5]:\n print(r[''document_title''], r[''page_numbers''], + r[''score''])\n"}' name: execute_code - id: call_arzz3ioj + id: call_00l27zrg index: 0 type: function - created: 1770373346 - id: chatcmpl-682 + created: 1771924503 + id: chatcmpl-643 model: gpt-oss object: chat.completion system_fingerprint: fp_ollama usage: - completion_tokens: 106 - prompt_tokens: 1749 - total_tokens: 1855 + completion_tokens: 110 + prompt_tokens: 1625 + total_tokens: 1735 status: code: 200 message: OK @@ -385,7 +357,47 @@ interactions: connection: - keep-alive content-length: - - '8709' + - '94' + content-type: + - application/json + host: + - localhost:11434 + method: POST + parsed_body: + encoding_format: base64 + input: + - quarterly report revenue + model: qwen3-embedding:4b + uri: http://localhost:11434/v1/embeddings + response: + headers: + content-type: + - application/json + transfer-encoding: + - chunked + parsed_body: + data: + - embedding: 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 + index: 0 + object: embedding + model: qwen3-embedding:4b + object: list + usage: + prompt_tokens: 5 + total_tokens: 5 + status: + code: 200 + message: OK +- request: + headers: + accept: + - application/json + accept-encoding: + - gzip, deflate, zstd + connection: + - keep-alive + content-length: + - '8336' content-type: - application/json host: @@ -398,32 +410,33 @@ interactions: IMPORTANT: You MUST use the `execute_code` tool to run Python code. The functions described below are ONLY available inside the execute_code tool - you cannot access them any other way. Always execute code to answer questions; do not just describe what code would do. - CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just use them directly: - - search("query") ✓ CORRECT + CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just call them with `await`: + - results = await search("query") ✓ CORRECT - from haiku.rag import search ✗ WRONG - will fail + - results = search("query") ✗ WRONG - must use await - You have access to a sandboxed Python environment with these haiku.rag functions (use them directly, no imports needed): + You have access to a sandboxed Python interpreter with these haiku.rag functions (use them directly with `await`, no imports needed): ## Available Functions - ### search(query, limit=10) -> list[dict] + ### await search(query, limit=10) -> list[dict] Search the knowledge base using hybrid search (vector + full-text). Returns list of dicts with keys: chunk_id, content, document_id, document_title, document_uri, score, page_numbers, headings - ### list_documents(limit=10, offset=0) -> list[dict] + ### await list_documents(limit=10, offset=0) -> list[dict] List available documents in the knowledge base. Returns list of dicts with keys: id, title, uri, created_at - ### get_document(id_or_title) -> str | None + ### await get_document(id_or_title) -> str | None Get the full text content of a document by ID, title, or URI. Returns the document content as a string, or None if not found. - ### get_docling_document(id_or_title) -> DoclingDocument | None - Get the structured DoclingDocument object for advanced analysis. - Returns a DoclingDocument object, or None if not found. - See "DoclingDocument API" section below for how to use it. + ### await get_chunk(chunk_id) -> dict | None + Get a specific chunk by its ID (from search results). + Returns dict with keys: chunk_id, content, document_id, document_title, headings, page_numbers, labels + Use this to retrieve full chunk details and metadata for citation. - ### llm(prompt) -> str + ### await llm(prompt) -> str Call an LLM directly with the given prompt. Returns the response as a string. Use this for classification, summarization, extraction, or any task where you already have the content and just need LLM reasoning. @@ -438,93 +451,64 @@ interactions: ``` Check if it exists with: `if 'documents' in dir(): ...` - ## Standard Library Modules - You can import any Python standard library module. + ## Available Python Features + + The interpreter supports: variables, arithmetic, strings, f-strings, lists, dicts, tuples, sets, loops, conditionals, comprehensions, functions, async/await, `map()`, `sorted()`/`.sort(key=...)`, try/except, and the `json` module. + + Not supported: imports (other than `json`), class definitions, generators/yield, match statements, decorators, `with` statements. + + For pattern matching or text extraction, use string methods (`str.split`, `str.find`, `str.startswith`, `in` operator) or the `llm()` function. ## Strategy Guide 1. **Explore First**: Start by listing documents or searching to understand what's available. Document names may differ from filenames (e.g., "tbmed593.pdf" might be stored as "TB MED 593" or similar). - 2. **If get_document returns None**: Use `list_documents()` to see actual document titles, or `search()` to find relevant content. + 2. **If get_document returns None**: Use `await list_documents()` to see actual document titles, or `await search()` to find relevant content. 3. **Iterative Refinement**: Run code, examine results, adjust your approach based on what you find. - 4. **Use print() Liberally**: The REPL captures stdout - print intermediate results to see what you're working with. - 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and use collections. - 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use llm(). - 7. **Cite Your Sources**: Track which documents/chunks informed your answer for citation. - - ## DoclingDocument API - - When you call `get_docling_document(id_or_title)`, you get a DoclingDocument object for structured document analysis. - - ### Properties - - `doc.texts` - List of all text items (paragraphs, headings, etc.) - - `doc.tables` - List of all tables - - `doc.pictures` - List of all pictures/figures - - `doc.name` - Document name - - ### Methods - - `doc.iterate_items(with_groups=False)` - Iterate all items with hierarchy level - Returns tuples of (item, level) where level is nesting depth - - `doc.export_to_markdown()` - Export entire document as markdown string - - ### Text Item Properties - - `item.text` - The text content - - `item.label` - Type: title, paragraph, section_header, list_item, etc. (lowercase enum values) - - `item.prov` - Provenance (page numbers, bounding boxes) - - ### Table Access - - `table.data.num_rows`, `table.data.num_cols` - Dimensions - - `table.data.table_cells` - List of TableCell objects - - `cell.text`, `cell.start_row_offset_idx`, `cell.start_col_offset_idx` - - ### Example Usage - ```python - doc = get_docling_document("My Document") - - # Get all headings - headings = [t.text for t in doc.texts if "header" in str(t.label)] - - # Iterate with structure - for item, level in doc.iterate_items(): - print(" " * level + item.text[:50]) - - # Extract table data - for table in doc.tables: - for cell in table.data.table_cells: - print(f"Row {cell.start_row_offset_idx}, Col {cell.start_col_offset_idx}: {cell.text}") - ``` + 4. **Use print() Liberally**: The sandbox captures stdout - print intermediate results to see what you're working with. + 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and data structures. + 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use `await llm()`. + 7. **Cite Your Sources**: Use get_chunk() to retrieve chunk metadata for citations. Track which documents/chunks informed your answer. ## Example Patterns ### Counting documents matching a condition ```python - docs = list_documents(limit=100) + docs = await list_documents(limit=100) count = 0 for doc in docs: - content = get_document(doc['id']) + content = await get_document(doc['id']) if content and 'keyword' in content.lower(): count += 1 print(f"Found in: {doc['title']}") print(f"Total: {count}") ``` - ### Aggregating data across documents + ### Extracting data with llm() ```python - import re numbers = [] - results = search("financial data", limit=20) + results = await search("financial data", limit=20) for r in results: - matches = re.findall(r'\$([\d,]+)', r['content']) - for m in matches: - numbers.append(int(m.replace(',', ''))) - print(f"Average: ${sum(numbers)/len(numbers):,.2f}") + extracted = await llm(f"Extract all dollar amounts from this text as a comma-separated list of numbers (no $ signs): {r['content']}") + for part in extracted.split(','): + part = part.strip().replace(',', '') + if part.isdigit(): + numbers.append(int(part)) + if numbers: + print(f"Average: {sum(numbers) / len(numbers)}") + ``` + + ### Using search results with get_chunk for citations + ```python + results = await search("safety requirements", limit=5) + for r in results: + chunk = await get_chunk(r['chunk_id']) + print(f"From '{chunk['document_title']}', page {chunk['page_numbers']}: {chunk['content'][:100]}") ``` ### Using llm() for classification ```python - # Get document content - content = get_document("Q1 Report") - # Use llm() to classify sentiment - sentiment = llm(f"Classify the sentiment as positive, negative, or mixed: {content}") + content = await get_document("Q1 Report") + sentiment = await llm(f"Classify the sentiment as positive, negative, or mixed: {content}") print(sentiment) ``` @@ -550,23 +534,23 @@ interactions: role: system - content: What is the total revenue across all quarterly reports? role: user - - content: |- - - Need to find quarterly reports, extract revenue amounts, sum. Likely documents titled like "Q1 Report", "Q2 Report"... Let's search "quarterly report revenue". - + - content: null + reasoning: We need to find quarterly reports documents and extract revenue figures. Likely documents include Q1, Q2, + etc. Search for "quarterly report revenue". role: assistant tool_calls: - function: - arguments: '{"code":"results = search(\"quarterly report revenue\", limit=20)\nprint(len(results))\nfor r in results[:5]:\n print(r[''document_title''], - r[''page_numbers''], r[''score''])\n"}' + arguments: '{"code":"# Search for quarterly report revenue\nresults = await search(\"quarterly report revenue\", + limit=20)\nprint(len(results))\nfor r in results[:5]:\n print(r[''document_title''], r[''page_numbers''], + r[''score''])\n"}' name: execute_code - id: call_arzz3ioj + id: call_00l27zrg type: function - - content: '{"code":"results = search(\"quarterly report revenue\", limit=20)\nprint(len(results))\nfor r in results[:5]:\n print(r[''document_title''], - r[''page_numbers''], r[''score''])\n","stdout":"3\nQ1 Report [] 0.03226645663380623\nQ2 Report [] 0.03226645663380623\nQ3 - Report [] 0.032258063554763794\n","stderr":"","success":true}' + - content: '{"code":"# Search for quarterly report revenue\nresults = await search(\"quarterly report revenue\", limit=20)\nprint(len(results))\nfor + r in results[:5]:\n print(r[''document_title''], r[''page_numbers''], r[''score''])\n","stdout":"3\nQ1 Report + [] 0.03226645663380623\nQ2 Report [] 0.03226645663380623\nQ3 Report [] 0.032258063554763794\n","stderr":"","success":true}' role: tool - tool_call_id: call_arzz3ioj + tool_call_id: call_00l27zrg model: gpt-oss reasoning_effort: low stream: false @@ -574,11 +558,10 @@ interactions: tools: - function: description: |- - Execute Python code in a Docker-sandboxed environment. + Execute Python code in a sandboxed interpreter. The code has access to haiku.rag functions (search, list_documents, - get_document, get_docling_document, llm) and any Python standard - library module. + get_document, get_chunk, llm). Use print() to output results. @@ -618,7 +601,7 @@ interactions: response: headers: content-length: - - '1307' + - '567' content-type: - application/json parsed_body: @@ -627,31 +610,24 @@ interactions: index: 0 message: content: '' - reasoning: We have 3 reports. Need to extract revenue. Let's fetch each document content and find revenue. Use regex - for "$...". + reasoning: Found Q1, Q2, Q3. Probably Q4 too. Search for Q4. role: assistant tool_calls: - function: - arguments: '{"code":"import re\nreports = [''Q1 Report'', ''Q2 Report'', ''Q3 Report'']\nrevenues = {}\nfor - rep in reports:\n content = get_document(rep)\n if content is None:\n # try by title\n for - doc in list_documents():\n if rep.lower() in doc[''title''].lower():\n content = - get_document(doc[''id''])\n break\n if content:\n # find patterns like \"Revenue: - $1,234,567\"\n matches = re.findall(r''Revenue[:\\s]*\\$([\\d,]+)'', content)\n amounts = [int(m.replace('','','''')) - for m in matches]\n if amounts:\n revenues[rep] = sum(amounts)\n else:\n revenues[rep] - = None\n else:\n revenues[rep] = None\nprint(revenues)\n"}' + arguments: '{"code":"results4 = await search(\"Q4 Report revenue\", limit=5)\nprint(results4)\n"}' name: execute_code - id: call_vuqzymvs + id: call_s2cf4xgs index: 0 type: function - created: 1770373351 - id: chatcmpl-118 + created: 1771924505 + id: chatcmpl-466 model: gpt-oss object: chat.completion system_fingerprint: fp_ollama usage: - completion_tokens: 245 - prompt_tokens: 1975 - total_tokens: 2220 + completion_tokens: 64 + prompt_tokens: 1858 + total_tokens: 1922 status: code: 200 message: OK @@ -664,292 +640,34 @@ interactions: connection: - keep-alive content-length: - - '10684' + - '87' content-type: - application/json host: - localhost:11434 method: POST parsed_body: - messages: - - content: |- - You are a Recursive Language Model (RLM) agent that solves complex research questions by writing and executing Python code. - - IMPORTANT: You MUST use the `execute_code` tool to run Python code. The functions described below are ONLY available inside the execute_code tool - you cannot access them any other way. Always execute code to answer questions; do not just describe what code would do. - - CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just use them directly: - - search("query") ✓ CORRECT - - from haiku.rag import search ✗ WRONG - will fail - - You have access to a sandboxed Python environment with these haiku.rag functions (use them directly, no imports needed): - - ## Available Functions - - ### search(query, limit=10) -> list[dict] - Search the knowledge base using hybrid search (vector + full-text). - Returns list of dicts with keys: chunk_id, content, document_id, document_title, document_uri, score, page_numbers, headings - - ### list_documents(limit=10, offset=0) -> list[dict] - List available documents in the knowledge base. - Returns list of dicts with keys: id, title, uri, created_at - - ### get_document(id_or_title) -> str | None - Get the full text content of a document by ID, title, or URI. - Returns the document content as a string, or None if not found. - - ### get_docling_document(id_or_title) -> DoclingDocument | None - Get the structured DoclingDocument object for advanced analysis. - Returns a DoclingDocument object, or None if not found. - See "DoclingDocument API" section below for how to use it. - - ### llm(prompt) -> str - Call an LLM directly with the given prompt. Returns the response as a string. - Use this for classification, summarization, extraction, or any task where you - already have the content and just need LLM reasoning. - - ## Pre-loaded Documents Variable - - If documents were pre-loaded for this session, a `documents` variable is available: - ```python - # documents is a list of dicts with keys: id, title, uri, content - for doc in documents: - print(doc['title'], len(doc['content'])) - ``` - Check if it exists with: `if 'documents' in dir(): ...` - - ## Standard Library Modules - You can import any Python standard library module. - - ## Strategy Guide - - 1. **Explore First**: Start by listing documents or searching to understand what's available. Document names may differ from filenames (e.g., "tbmed593.pdf" might be stored as "TB MED 593" or similar). - 2. **If get_document returns None**: Use `list_documents()` to see actual document titles, or `search()` to find relevant content. - 3. **Iterative Refinement**: Run code, examine results, adjust your approach based on what you find. - 4. **Use print() Liberally**: The REPL captures stdout - print intermediate results to see what you're working with. - 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and use collections. - 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use llm(). - 7. **Cite Your Sources**: Track which documents/chunks informed your answer for citation. - - ## DoclingDocument API - - When you call `get_docling_document(id_or_title)`, you get a DoclingDocument object for structured document analysis. - - ### Properties - - `doc.texts` - List of all text items (paragraphs, headings, etc.) - - `doc.tables` - List of all tables - - `doc.pictures` - List of all pictures/figures - - `doc.name` - Document name - - ### Methods - - `doc.iterate_items(with_groups=False)` - Iterate all items with hierarchy level - Returns tuples of (item, level) where level is nesting depth - - `doc.export_to_markdown()` - Export entire document as markdown string - - ### Text Item Properties - - `item.text` - The text content - - `item.label` - Type: title, paragraph, section_header, list_item, etc. (lowercase enum values) - - `item.prov` - Provenance (page numbers, bounding boxes) - - ### Table Access - - `table.data.num_rows`, `table.data.num_cols` - Dimensions - - `table.data.table_cells` - List of TableCell objects - - `cell.text`, `cell.start_row_offset_idx`, `cell.start_col_offset_idx` - - ### Example Usage - ```python - doc = get_docling_document("My Document") - - # Get all headings - headings = [t.text for t in doc.texts if "header" in str(t.label)] - - # Iterate with structure - for item, level in doc.iterate_items(): - print(" " * level + item.text[:50]) - - # Extract table data - for table in doc.tables: - for cell in table.data.table_cells: - print(f"Row {cell.start_row_offset_idx}, Col {cell.start_col_offset_idx}: {cell.text}") - ``` - - ## Example Patterns - - ### Counting documents matching a condition - ```python - docs = list_documents(limit=100) - count = 0 - for doc in docs: - content = get_document(doc['id']) - if content and 'keyword' in content.lower(): - count += 1 - print(f"Found in: {doc['title']}") - print(f"Total: {count}") - ``` - - ### Aggregating data across documents - ```python - import re - numbers = [] - results = search("financial data", limit=20) - for r in results: - matches = re.findall(r'\$([\d,]+)', r['content']) - for m in matches: - numbers.append(int(m.replace(',', ''))) - print(f"Average: ${sum(numbers)/len(numbers):,.2f}") - ``` - - ### Using llm() for classification - ```python - # Get document content - content = get_document("Q1 Report") - # Use llm() to classify sentiment - sentiment = llm(f"Classify the sentiment as positive, negative, or mixed: {content}") - print(sentiment) - ``` - - ## Workflow - - 1. **ALWAYS start by using execute_code** to explore the knowledge base - 2. Run multiple code blocks as needed to gather information - 3. After collecting data, provide your final answer - - ## Output Format - - CRITICAL: Your final response MUST be valid JSON matching this exact schema: - ```json - {"answer": "Your complete answer here as a string", "program": "Your final consolidated program here as a string"} - ``` - - - `answer`: A clear answer to the user's question with key findings and references to specific documents/chunks. - - `program`: A single, self-contained Python program that produces the answer. Consolidate your exploratory code executions into one clean script. - - Do NOT return arbitrary JSON structures. Always use the exact format: {"answer": "...", "program": "..."} - - CRITICAL: You MUST call execute_code at least once before providing your answer. Never give up without trying to execute code first. - role: system - - content: What is the total revenue across all quarterly reports? - role: user - - content: |- - - Need to find quarterly reports, extract revenue amounts, sum. Likely documents titled like "Q1 Report", "Q2 Report"... Let's search "quarterly report revenue". - - role: assistant - tool_calls: - - function: - arguments: '{"code":"results = search(\"quarterly report revenue\", limit=20)\nprint(len(results))\nfor r in results[:5]:\n print(r[''document_title''], - r[''page_numbers''], r[''score''])\n"}' - name: execute_code - id: call_arzz3ioj - type: function - - content: '{"code":"results = search(\"quarterly report revenue\", limit=20)\nprint(len(results))\nfor r in results[:5]:\n print(r[''document_title''], - r[''page_numbers''], r[''score''])\n","stdout":"3\nQ1 Report [] 0.03226645663380623\nQ2 Report [] 0.03226645663380623\nQ3 - Report [] 0.032258063554763794\n","stderr":"","success":true}' - role: tool - tool_call_id: call_arzz3ioj - - content: |- - - We have 3 reports. Need to extract revenue. Let's fetch each document content and find revenue. Use regex for "$...". - - role: assistant - tool_calls: - - function: - arguments: '{"code":"import re\nreports = [''Q1 Report'', ''Q2 Report'', ''Q3 Report'']\nrevenues = {}\nfor rep - in reports:\n content = get_document(rep)\n if content is None:\n # try by title\n for doc - in list_documents():\n if rep.lower() in doc[''title''].lower():\n content = get_document(doc[''id''])\n break\n if - content:\n # find patterns like \"Revenue: $1,234,567\"\n matches = re.findall(r''Revenue[:\\s]*\\$([\\d,]+)'', - content)\n amounts = [int(m.replace('','','''')) for m in matches]\n if amounts:\n revenues[rep] - = sum(amounts)\n else:\n revenues[rep] = None\n else:\n revenues[rep] = None\nprint(revenues)\n"}' - name: execute_code - id: call_vuqzymvs - type: function - - content: '{"code":"import re\nreports = [''Q1 Report'', ''Q2 Report'', ''Q3 Report'']\nrevenues = {}\nfor rep in reports:\n content - = get_document(rep)\n if content is None:\n # try by title\n for doc in list_documents():\n if - rep.lower() in doc[''title''].lower():\n content = get_document(doc[''id''])\n break\n if - content:\n # find patterns like \"Revenue: $1,234,567\"\n matches = re.findall(r''Revenue[:\\s]*\\$([\\d,]+)'', - content)\n amounts = [int(m.replace('','','''')) for m in matches]\n if amounts:\n revenues[rep] - = sum(amounts)\n else:\n revenues[rep] = None\n else:\n revenues[rep] = None\nprint(revenues)\n","stdout":"{''Q1 - Report'': None, ''Q2 Report'': None, ''Q3 Report'': None}\n","stderr":"","success":true}' - role: tool - tool_call_id: call_vuqzymvs - model: gpt-oss - reasoning_effort: low - stream: false - tool_choice: auto - tools: - - function: - description: |- - Execute Python code in a Docker-sandboxed environment. - - The code has access to haiku.rag functions (search, list_documents, - get_document, get_docling_document, llm) and any Python standard - library module. - - Use print() to output results. - - Structured result with success status, stdout, and stderr. - - name: execute_code - parameters: - additionalProperties: false - properties: - code: - description: Python code to execute. - type: string - required: - - code - type: object - strict: true - type: function - - function: - description: Result from RLM agent execution. - name: final_result - parameters: - additionalProperties: false - properties: - answer: - description: The answer to the user's question - type: string - program: - description: The final consolidated program - type: string - required: - - answer - - program - type: object - strict: true - type: function - uri: http://localhost:11434/v1/chat/completions + encoding_format: base64 + input: + - Q4 Report revenue + model: qwen3-embedding:4b + uri: http://localhost:11434/v1/embeddings response: headers: - content-length: - - '626' content-type: - application/json + transfer-encoding: + - chunked parsed_body: - choices: - - finish_reason: tool_calls + data: + - embedding: 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 index: 0 - message: - content: |- - - No content or revenue not matched. Let's try list documents to see actual names and perhaps extract via pages. - role: assistant - tool_calls: - - function: - arguments: '{"code":"docs = list_documents(limit=50)\nfor d in docs:\n print(d[''title''])\n"}' - name: execute_code - id: call_z0mgsixf - index: 0 - type: function - created: 1770373353 - id: chatcmpl-25 - model: gpt-oss - object: chat.completion - system_fingerprint: fp_ollama + object: embedding + model: qwen3-embedding:4b + object: list usage: - completion_tokens: 70 - prompt_tokens: 2468 - total_tokens: 2538 + prompt_tokens: 5 + total_tokens: 5 status: code: 200 message: OK @@ -962,7 +680,7 @@ interactions: connection: - keep-alive content-length: - - '11261' + - '9694' content-type: - application/json host: @@ -975,32 +693,33 @@ interactions: IMPORTANT: You MUST use the `execute_code` tool to run Python code. The functions described below are ONLY available inside the execute_code tool - you cannot access them any other way. Always execute code to answer questions; do not just describe what code would do. - CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just use them directly: - - search("query") ✓ CORRECT + CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just call them with `await`: + - results = await search("query") ✓ CORRECT - from haiku.rag import search ✗ WRONG - will fail + - results = search("query") ✗ WRONG - must use await - You have access to a sandboxed Python environment with these haiku.rag functions (use them directly, no imports needed): + You have access to a sandboxed Python interpreter with these haiku.rag functions (use them directly with `await`, no imports needed): ## Available Functions - ### search(query, limit=10) -> list[dict] + ### await search(query, limit=10) -> list[dict] Search the knowledge base using hybrid search (vector + full-text). Returns list of dicts with keys: chunk_id, content, document_id, document_title, document_uri, score, page_numbers, headings - ### list_documents(limit=10, offset=0) -> list[dict] + ### await list_documents(limit=10, offset=0) -> list[dict] List available documents in the knowledge base. Returns list of dicts with keys: id, title, uri, created_at - ### get_document(id_or_title) -> str | None + ### await get_document(id_or_title) -> str | None Get the full text content of a document by ID, title, or URI. Returns the document content as a string, or None if not found. - ### get_docling_document(id_or_title) -> DoclingDocument | None - Get the structured DoclingDocument object for advanced analysis. - Returns a DoclingDocument object, or None if not found. - See "DoclingDocument API" section below for how to use it. + ### await get_chunk(chunk_id) -> dict | None + Get a specific chunk by its ID (from search results). + Returns dict with keys: chunk_id, content, document_id, document_title, headings, page_numbers, labels + Use this to retrieve full chunk details and metadata for citation. - ### llm(prompt) -> str + ### await llm(prompt) -> str Call an LLM directly with the given prompt. Returns the response as a string. Use this for classification, summarization, extraction, or any task where you already have the content and just need LLM reasoning. @@ -1015,93 +734,64 @@ interactions: ``` Check if it exists with: `if 'documents' in dir(): ...` - ## Standard Library Modules - You can import any Python standard library module. + ## Available Python Features + + The interpreter supports: variables, arithmetic, strings, f-strings, lists, dicts, tuples, sets, loops, conditionals, comprehensions, functions, async/await, `map()`, `sorted()`/`.sort(key=...)`, try/except, and the `json` module. + + Not supported: imports (other than `json`), class definitions, generators/yield, match statements, decorators, `with` statements. + + For pattern matching or text extraction, use string methods (`str.split`, `str.find`, `str.startswith`, `in` operator) or the `llm()` function. ## Strategy Guide 1. **Explore First**: Start by listing documents or searching to understand what's available. Document names may differ from filenames (e.g., "tbmed593.pdf" might be stored as "TB MED 593" or similar). - 2. **If get_document returns None**: Use `list_documents()` to see actual document titles, or `search()` to find relevant content. + 2. **If get_document returns None**: Use `await list_documents()` to see actual document titles, or `await search()` to find relevant content. 3. **Iterative Refinement**: Run code, examine results, adjust your approach based on what you find. - 4. **Use print() Liberally**: The REPL captures stdout - print intermediate results to see what you're working with. - 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and use collections. - 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use llm(). - 7. **Cite Your Sources**: Track which documents/chunks informed your answer for citation. - - ## DoclingDocument API - - When you call `get_docling_document(id_or_title)`, you get a DoclingDocument object for structured document analysis. - - ### Properties - - `doc.texts` - List of all text items (paragraphs, headings, etc.) - - `doc.tables` - List of all tables - - `doc.pictures` - List of all pictures/figures - - `doc.name` - Document name - - ### Methods - - `doc.iterate_items(with_groups=False)` - Iterate all items with hierarchy level - Returns tuples of (item, level) where level is nesting depth - - `doc.export_to_markdown()` - Export entire document as markdown string - - ### Text Item Properties - - `item.text` - The text content - - `item.label` - Type: title, paragraph, section_header, list_item, etc. (lowercase enum values) - - `item.prov` - Provenance (page numbers, bounding boxes) - - ### Table Access - - `table.data.num_rows`, `table.data.num_cols` - Dimensions - - `table.data.table_cells` - List of TableCell objects - - `cell.text`, `cell.start_row_offset_idx`, `cell.start_col_offset_idx` - - ### Example Usage - ```python - doc = get_docling_document("My Document") - - # Get all headings - headings = [t.text for t in doc.texts if "header" in str(t.label)] - - # Iterate with structure - for item, level in doc.iterate_items(): - print(" " * level + item.text[:50]) - - # Extract table data - for table in doc.tables: - for cell in table.data.table_cells: - print(f"Row {cell.start_row_offset_idx}, Col {cell.start_col_offset_idx}: {cell.text}") - ``` + 4. **Use print() Liberally**: The sandbox captures stdout - print intermediate results to see what you're working with. + 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and data structures. + 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use `await llm()`. + 7. **Cite Your Sources**: Use get_chunk() to retrieve chunk metadata for citations. Track which documents/chunks informed your answer. ## Example Patterns ### Counting documents matching a condition ```python - docs = list_documents(limit=100) + docs = await list_documents(limit=100) count = 0 for doc in docs: - content = get_document(doc['id']) + content = await get_document(doc['id']) if content and 'keyword' in content.lower(): count += 1 print(f"Found in: {doc['title']}") print(f"Total: {count}") ``` - ### Aggregating data across documents + ### Extracting data with llm() ```python - import re numbers = [] - results = search("financial data", limit=20) + results = await search("financial data", limit=20) for r in results: - matches = re.findall(r'\$([\d,]+)', r['content']) - for m in matches: - numbers.append(int(m.replace(',', ''))) - print(f"Average: ${sum(numbers)/len(numbers):,.2f}") + extracted = await llm(f"Extract all dollar amounts from this text as a comma-separated list of numbers (no $ signs): {r['content']}") + for part in extracted.split(','): + part = part.strip().replace(',', '') + if part.isdigit(): + numbers.append(int(part)) + if numbers: + print(f"Average: {sum(numbers) / len(numbers)}") + ``` + + ### Using search results with get_chunk for citations + ```python + results = await search("safety requirements", limit=5) + for r in results: + chunk = await get_chunk(r['chunk_id']) + print(f"From '{chunk['document_title']}', page {chunk['page_numbers']}: {chunk['content'][:100]}") ``` ### Using llm() for classification ```python - # Get document content - content = get_document("Q1 Report") - # Use llm() to classify sentiment - sentiment = llm(f"Classify the sentiment as positive, negative, or mixed: {content}") + content = await get_document("Q1 Report") + sentiment = await llm(f"Classify the sentiment as positive, negative, or mixed: {content}") print(sentiment) ``` @@ -1127,62 +817,43 @@ interactions: role: system - content: What is the total revenue across all quarterly reports? role: user - - content: |- - - Need to find quarterly reports, extract revenue amounts, sum. Likely documents titled like "Q1 Report", "Q2 Report"... Let's search "quarterly report revenue". - + - content: null + reasoning: We need to find quarterly reports documents and extract revenue figures. Likely documents include Q1, Q2, + etc. Search for "quarterly report revenue". role: assistant tool_calls: - function: - arguments: '{"code":"results = search(\"quarterly report revenue\", limit=20)\nprint(len(results))\nfor r in results[:5]:\n print(r[''document_title''], - r[''page_numbers''], r[''score''])\n"}' + arguments: '{"code":"# Search for quarterly report revenue\nresults = await search(\"quarterly report revenue\", + limit=20)\nprint(len(results))\nfor r in results[:5]:\n print(r[''document_title''], r[''page_numbers''], + r[''score''])\n"}' name: execute_code - id: call_arzz3ioj + id: call_00l27zrg type: function - - content: '{"code":"results = search(\"quarterly report revenue\", limit=20)\nprint(len(results))\nfor r in results[:5]:\n print(r[''document_title''], - r[''page_numbers''], r[''score''])\n","stdout":"3\nQ1 Report [] 0.03226645663380623\nQ2 Report [] 0.03226645663380623\nQ3 - Report [] 0.032258063554763794\n","stderr":"","success":true}' + - content: '{"code":"# Search for quarterly report revenue\nresults = await search(\"quarterly report revenue\", limit=20)\nprint(len(results))\nfor + r in results[:5]:\n print(r[''document_title''], r[''page_numbers''], r[''score''])\n","stdout":"3\nQ1 Report + [] 0.03226645663380623\nQ2 Report [] 0.03226645663380623\nQ3 Report [] 0.032258063554763794\n","stderr":"","success":true}' role: tool - tool_call_id: call_arzz3ioj - - content: |- - - We have 3 reports. Need to extract revenue. Let's fetch each document content and find revenue. Use regex for "$...". - + tool_call_id: call_00l27zrg + - content: null + reasoning: Found Q1, Q2, Q3. Probably Q4 too. Search for Q4. role: assistant tool_calls: - function: - arguments: '{"code":"import re\nreports = [''Q1 Report'', ''Q2 Report'', ''Q3 Report'']\nrevenues = {}\nfor rep - in reports:\n content = get_document(rep)\n if content is None:\n # try by title\n for doc - in list_documents():\n if rep.lower() in doc[''title''].lower():\n content = get_document(doc[''id''])\n break\n if - content:\n # find patterns like \"Revenue: $1,234,567\"\n matches = re.findall(r''Revenue[:\\s]*\\$([\\d,]+)'', - content)\n amounts = [int(m.replace('','','''')) for m in matches]\n if amounts:\n revenues[rep] - = sum(amounts)\n else:\n revenues[rep] = None\n else:\n revenues[rep] = None\nprint(revenues)\n"}' + arguments: '{"code":"results4 = await search(\"Q4 Report revenue\", limit=5)\nprint(results4)\n"}' name: execute_code - id: call_vuqzymvs + id: call_s2cf4xgs type: function - - content: '{"code":"import re\nreports = [''Q1 Report'', ''Q2 Report'', ''Q3 Report'']\nrevenues = {}\nfor rep in reports:\n content - = get_document(rep)\n if content is None:\n # try by title\n for doc in list_documents():\n if - rep.lower() in doc[''title''].lower():\n content = get_document(doc[''id''])\n break\n if - content:\n # find patterns like \"Revenue: $1,234,567\"\n matches = re.findall(r''Revenue[:\\s]*\\$([\\d,]+)'', - content)\n amounts = [int(m.replace('','','''')) for m in matches]\n if amounts:\n revenues[rep] - = sum(amounts)\n else:\n revenues[rep] = None\n else:\n revenues[rep] = None\nprint(revenues)\n","stdout":"{''Q1 - Report'': None, ''Q2 Report'': None, ''Q3 Report'': None}\n","stderr":"","success":true}' + - content: '{"code":"results4 = await search(\"Q4 Report revenue\", limit=5)\nprint(results4)\n","stdout":"[{''chunk_id'': + ''fb347b1c-9bf0-455b-8535-3484f9578731'', ''content'': ''Sales report Q1: Revenue was $100,000.'', ''document_id'': + ''7c76fa0e-4d74-49d1-88a6-908391e63fe5'', ''document_title'': ''Q1 Report'', ''document_uri'': None, ''score'': + 0.03226645663380623, ''page_numbers'': [], ''headings'': None}, {''chunk_id'': ''50f60a6f-53e4-4850-8d5b-a3c00d4375c6'', + ''content'': ''Sales report Q2: Revenue was $150,000.'', ''document_id'': ''792e21c1-2736-4fa8-ba8e-0629b1cb5a2f'', + ''document_title'': ''Q2 Report'', ''document_uri'': None, ''score'': 0.03226645663380623, ''page_numbers'': [], + ''headings'': None}, {''chunk_id'': ''70233263-be1a-414e-8841-c9509076530e'', ''content'': ''Sales report Q3: Revenue + was $200,000.'', ''document_id'': ''ab42afcc-0fe0-4fdc-8c26-656d747c950e'', ''document_title'': ''Q3 Report'', ''document_uri'': + None, ''score'': 0.032258063554763794, ''page_numbers'': [], ''headings'': None}]\n","stderr":"","success":true}' role: tool - tool_call_id: call_vuqzymvs - - content: |2- - - No content or revenue not matched. Let's try list documents to see actual names and perhaps extract via pages. - role: assistant - tool_calls: - - function: - arguments: '{"code":"docs = list_documents(limit=50)\nfor d in docs:\n print(d[''title''])\n"}' - name: execute_code - id: call_z0mgsixf - type: function - - content: '{"code":"docs = list_documents(limit=50)\nfor d in docs:\n print(d[''title''])\n","stdout":"Q1 Report\nQ2 - Report\nQ3 Report\n","stderr":"","success":true}' - role: tool - tool_call_id: call_z0mgsixf + tool_call_id: call_s2cf4xgs model: gpt-oss reasoning_effort: low stream: false @@ -1190,11 +861,10 @@ interactions: tools: - function: description: |- - Execute Python code in a Docker-sandboxed environment. + Execute Python code in a sandboxed interpreter. The code has access to haiku.rag functions (search, list_documents, - get_document, get_docling_document, llm) and any Python standard - library module. + get_document, get_chunk, llm). Use print() to output results. @@ -1234,1436 +904,7 @@ interactions: response: headers: content-length: - - '859' - content-type: - - application/json - parsed_body: - choices: - - finish_reason: tool_calls - index: 0 - message: - content: '' - reasoning: Maybe revenue is in tables. Use docling. - role: assistant - tool_calls: - - function: - arguments: '{"code":"doc = get_docling_document(''Q1 Report'')\nprint(doc.name)\n# print headings\nheadings - = [t.text for t in doc.texts if ''header'' in str(t.label)]\nprint(''headings:'', headlines:=headings[:5])\n# - find tables\nprint(f\"Tables: {len(doc.tables)}\")\nfor i,table in enumerate(doc.tables):\n print(f\"Table - {i} rows {table.data.num_rows} cols {table.data.num_cols}\")\n"}' - name: execute_code - id: call_acbn2mdz - index: 0 - type: function - created: 1770373356 - id: chatcmpl-268 - model: gpt-oss - object: chat.completion - system_fingerprint: fp_ollama - usage: - completion_tokens: 145 - prompt_tokens: 2595 - total_tokens: 2740 - status: - code: 200 - message: OK -- request: - headers: - accept: - - application/json - accept-encoding: - - gzip, deflate, zstd - connection: - - keep-alive - content-length: - - '12396' - content-type: - - application/json - host: - - localhost:11434 - method: POST - parsed_body: - messages: - - content: |- - You are a Recursive Language Model (RLM) agent that solves complex research questions by writing and executing Python code. - - IMPORTANT: You MUST use the `execute_code` tool to run Python code. The functions described below are ONLY available inside the execute_code tool - you cannot access them any other way. Always execute code to answer questions; do not just describe what code would do. - - CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just use them directly: - - search("query") ✓ CORRECT - - from haiku.rag import search ✗ WRONG - will fail - - You have access to a sandboxed Python environment with these haiku.rag functions (use them directly, no imports needed): - - ## Available Functions - - ### search(query, limit=10) -> list[dict] - Search the knowledge base using hybrid search (vector + full-text). - Returns list of dicts with keys: chunk_id, content, document_id, document_title, document_uri, score, page_numbers, headings - - ### list_documents(limit=10, offset=0) -> list[dict] - List available documents in the knowledge base. - Returns list of dicts with keys: id, title, uri, created_at - - ### get_document(id_or_title) -> str | None - Get the full text content of a document by ID, title, or URI. - Returns the document content as a string, or None if not found. - - ### get_docling_document(id_or_title) -> DoclingDocument | None - Get the structured DoclingDocument object for advanced analysis. - Returns a DoclingDocument object, or None if not found. - See "DoclingDocument API" section below for how to use it. - - ### llm(prompt) -> str - Call an LLM directly with the given prompt. Returns the response as a string. - Use this for classification, summarization, extraction, or any task where you - already have the content and just need LLM reasoning. - - ## Pre-loaded Documents Variable - - If documents were pre-loaded for this session, a `documents` variable is available: - ```python - # documents is a list of dicts with keys: id, title, uri, content - for doc in documents: - print(doc['title'], len(doc['content'])) - ``` - Check if it exists with: `if 'documents' in dir(): ...` - - ## Standard Library Modules - You can import any Python standard library module. - - ## Strategy Guide - - 1. **Explore First**: Start by listing documents or searching to understand what's available. Document names may differ from filenames (e.g., "tbmed593.pdf" might be stored as "TB MED 593" or similar). - 2. **If get_document returns None**: Use `list_documents()` to see actual document titles, or `search()` to find relevant content. - 3. **Iterative Refinement**: Run code, examine results, adjust your approach based on what you find. - 4. **Use print() Liberally**: The REPL captures stdout - print intermediate results to see what you're working with. - 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and use collections. - 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use llm(). - 7. **Cite Your Sources**: Track which documents/chunks informed your answer for citation. - - ## DoclingDocument API - - When you call `get_docling_document(id_or_title)`, you get a DoclingDocument object for structured document analysis. - - ### Properties - - `doc.texts` - List of all text items (paragraphs, headings, etc.) - - `doc.tables` - List of all tables - - `doc.pictures` - List of all pictures/figures - - `doc.name` - Document name - - ### Methods - - `doc.iterate_items(with_groups=False)` - Iterate all items with hierarchy level - Returns tuples of (item, level) where level is nesting depth - - `doc.export_to_markdown()` - Export entire document as markdown string - - ### Text Item Properties - - `item.text` - The text content - - `item.label` - Type: title, paragraph, section_header, list_item, etc. (lowercase enum values) - - `item.prov` - Provenance (page numbers, bounding boxes) - - ### Table Access - - `table.data.num_rows`, `table.data.num_cols` - Dimensions - - `table.data.table_cells` - List of TableCell objects - - `cell.text`, `cell.start_row_offset_idx`, `cell.start_col_offset_idx` - - ### Example Usage - ```python - doc = get_docling_document("My Document") - - # Get all headings - headings = [t.text for t in doc.texts if "header" in str(t.label)] - - # Iterate with structure - for item, level in doc.iterate_items(): - print(" " * level + item.text[:50]) - - # Extract table data - for table in doc.tables: - for cell in table.data.table_cells: - print(f"Row {cell.start_row_offset_idx}, Col {cell.start_col_offset_idx}: {cell.text}") - ``` - - ## Example Patterns - - ### Counting documents matching a condition - ```python - docs = list_documents(limit=100) - count = 0 - for doc in docs: - content = get_document(doc['id']) - if content and 'keyword' in content.lower(): - count += 1 - print(f"Found in: {doc['title']}") - print(f"Total: {count}") - ``` - - ### Aggregating data across documents - ```python - import re - numbers = [] - results = search("financial data", limit=20) - for r in results: - matches = re.findall(r'\$([\d,]+)', r['content']) - for m in matches: - numbers.append(int(m.replace(',', ''))) - print(f"Average: ${sum(numbers)/len(numbers):,.2f}") - ``` - - ### Using llm() for classification - ```python - # Get document content - content = get_document("Q1 Report") - # Use llm() to classify sentiment - sentiment = llm(f"Classify the sentiment as positive, negative, or mixed: {content}") - print(sentiment) - ``` - - ## Workflow - - 1. **ALWAYS start by using execute_code** to explore the knowledge base - 2. Run multiple code blocks as needed to gather information - 3. After collecting data, provide your final answer - - ## Output Format - - CRITICAL: Your final response MUST be valid JSON matching this exact schema: - ```json - {"answer": "Your complete answer here as a string", "program": "Your final consolidated program here as a string"} - ``` - - - `answer`: A clear answer to the user's question with key findings and references to specific documents/chunks. - - `program`: A single, self-contained Python program that produces the answer. Consolidate your exploratory code executions into one clean script. - - Do NOT return arbitrary JSON structures. Always use the exact format: {"answer": "...", "program": "..."} - - CRITICAL: You MUST call execute_code at least once before providing your answer. Never give up without trying to execute code first. - role: system - - content: What is the total revenue across all quarterly reports? - role: user - - content: |- - - Need to find quarterly reports, extract revenue amounts, sum. Likely documents titled like "Q1 Report", "Q2 Report"... Let's search "quarterly report revenue". - - role: assistant - tool_calls: - - function: - arguments: '{"code":"results = search(\"quarterly report revenue\", limit=20)\nprint(len(results))\nfor r in results[:5]:\n print(r[''document_title''], - r[''page_numbers''], r[''score''])\n"}' - name: execute_code - id: call_arzz3ioj - type: function - - content: '{"code":"results = search(\"quarterly report revenue\", limit=20)\nprint(len(results))\nfor r in results[:5]:\n print(r[''document_title''], - r[''page_numbers''], r[''score''])\n","stdout":"3\nQ1 Report [] 0.03226645663380623\nQ2 Report [] 0.03226645663380623\nQ3 - Report [] 0.032258063554763794\n","stderr":"","success":true}' - role: tool - tool_call_id: call_arzz3ioj - - content: |- - - We have 3 reports. Need to extract revenue. Let's fetch each document content and find revenue. Use regex for "$...". - - role: assistant - tool_calls: - - function: - arguments: '{"code":"import re\nreports = [''Q1 Report'', ''Q2 Report'', ''Q3 Report'']\nrevenues = {}\nfor rep - in reports:\n content = get_document(rep)\n if content is None:\n # try by title\n for doc - in list_documents():\n if rep.lower() in doc[''title''].lower():\n content = get_document(doc[''id''])\n break\n if - content:\n # find patterns like \"Revenue: $1,234,567\"\n matches = re.findall(r''Revenue[:\\s]*\\$([\\d,]+)'', - content)\n amounts = [int(m.replace('','','''')) for m in matches]\n if amounts:\n revenues[rep] - = sum(amounts)\n else:\n revenues[rep] = None\n else:\n revenues[rep] = None\nprint(revenues)\n"}' - name: execute_code - id: call_vuqzymvs - type: function - - content: '{"code":"import re\nreports = [''Q1 Report'', ''Q2 Report'', ''Q3 Report'']\nrevenues = {}\nfor rep in reports:\n content - = get_document(rep)\n if content is None:\n # try by title\n for doc in list_documents():\n if - rep.lower() in doc[''title''].lower():\n content = get_document(doc[''id''])\n break\n if - content:\n # find patterns like \"Revenue: $1,234,567\"\n matches = re.findall(r''Revenue[:\\s]*\\$([\\d,]+)'', - content)\n amounts = [int(m.replace('','','''')) for m in matches]\n if amounts:\n revenues[rep] - = sum(amounts)\n else:\n revenues[rep] = None\n else:\n revenues[rep] = None\nprint(revenues)\n","stdout":"{''Q1 - Report'': None, ''Q2 Report'': None, ''Q3 Report'': None}\n","stderr":"","success":true}' - role: tool - tool_call_id: call_vuqzymvs - - content: |2- - - No content or revenue not matched. Let's try list documents to see actual names and perhaps extract via pages. - role: assistant - tool_calls: - - function: - arguments: '{"code":"docs = list_documents(limit=50)\nfor d in docs:\n print(d[''title''])\n"}' - name: execute_code - id: call_z0mgsixf - type: function - - content: '{"code":"docs = list_documents(limit=50)\nfor d in docs:\n print(d[''title''])\n","stdout":"Q1 Report\nQ2 - Report\nQ3 Report\n","stderr":"","success":true}' - role: tool - tool_call_id: call_z0mgsixf - - content: |- - - Maybe revenue is in tables. Use docling. - - role: assistant - tool_calls: - - function: - arguments: '{"code":"doc = get_docling_document(''Q1 Report'')\nprint(doc.name)\n# print headings\nheadings = - [t.text for t in doc.texts if ''header'' in str(t.label)]\nprint(''headings:'', headlines:=headings[:5])\n# - find tables\nprint(f\"Tables: {len(doc.tables)}\")\nfor i,table in enumerate(doc.tables):\n print(f\"Table - {i} rows {table.data.num_rows} cols {table.data.num_cols}\")\n"}' - name: execute_code - id: call_acbn2mdz - type: function - - content: '{"code":"doc = get_docling_document(''Q1 Report'')\nprint(doc.name)\n# print headings\nheadings = [t.text - for t in doc.texts if ''header'' in str(t.label)]\nprint(''headings:'', headlines:=headings[:5])\n# find tables\nprint(f\"Tables: - {len(doc.tables)}\")\nfor i,table in enumerate(doc.tables):\n print(f\"Table {i} rows {table.data.num_rows} cols - {table.data.num_cols}\")\n","stdout":"content\nheadings: []\nTables: 0\n","stderr":"","success":true}' - role: tool - tool_call_id: call_acbn2mdz - model: gpt-oss - reasoning_effort: low - stream: false - tool_choice: auto - tools: - - function: - description: |- - Execute Python code in a Docker-sandboxed environment. - - The code has access to haiku.rag functions (search, list_documents, - get_document, get_docling_document, llm) and any Python standard - library module. - - Use print() to output results. - - Structured result with success status, stdout, and stderr. - - name: execute_code - parameters: - additionalProperties: false - properties: - code: - description: Python code to execute. - type: string - required: - - code - type: object - strict: true - type: function - - function: - description: Result from RLM agent execution. - name: final_result - parameters: - additionalProperties: false - properties: - answer: - description: The answer to the user's question - type: string - program: - description: The final consolidated program - type: string - required: - - answer - - program - type: object - strict: true - type: function - uri: http://localhost:11434/v1/chat/completions - response: - headers: - content-length: - - '643' - content-type: - - application/json - parsed_body: - choices: - - finish_reason: tool_calls - index: 0 - message: - content: '' - reasoning: No tables. Probably revenue mentioned in paragraphs. Let's fetch entire content and search for "$". - role: assistant - tool_calls: - - function: - arguments: '{"code":"content = get_document(''Q1 Report'')\nprint(content[:500])\nprint(''...'')\nprint(content.count(''$''))\n"}' - name: execute_code - id: call_y1fiwuj8 - index: 0 - type: function - created: 1770373358 - id: chatcmpl-409 - model: gpt-oss - object: chat.completion - system_fingerprint: fp_ollama - usage: - completion_tokens: 72 - prompt_tokens: 2895 - total_tokens: 2967 - status: - code: 200 - message: OK -- request: - headers: - accept: - - application/json - accept-encoding: - - gzip, deflate, zstd - connection: - - keep-alive - content-length: - - '13054' - content-type: - - application/json - host: - - localhost:11434 - method: POST - parsed_body: - messages: - - content: |- - You are a Recursive Language Model (RLM) agent that solves complex research questions by writing and executing Python code. - - IMPORTANT: You MUST use the `execute_code` tool to run Python code. The functions described below are ONLY available inside the execute_code tool - you cannot access them any other way. Always execute code to answer questions; do not just describe what code would do. - - CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just use them directly: - - search("query") ✓ CORRECT - - from haiku.rag import search ✗ WRONG - will fail - - You have access to a sandboxed Python environment with these haiku.rag functions (use them directly, no imports needed): - - ## Available Functions - - ### search(query, limit=10) -> list[dict] - Search the knowledge base using hybrid search (vector + full-text). - Returns list of dicts with keys: chunk_id, content, document_id, document_title, document_uri, score, page_numbers, headings - - ### list_documents(limit=10, offset=0) -> list[dict] - List available documents in the knowledge base. - Returns list of dicts with keys: id, title, uri, created_at - - ### get_document(id_or_title) -> str | None - Get the full text content of a document by ID, title, or URI. - Returns the document content as a string, or None if not found. - - ### get_docling_document(id_or_title) -> DoclingDocument | None - Get the structured DoclingDocument object for advanced analysis. - Returns a DoclingDocument object, or None if not found. - See "DoclingDocument API" section below for how to use it. - - ### llm(prompt) -> str - Call an LLM directly with the given prompt. Returns the response as a string. - Use this for classification, summarization, extraction, or any task where you - already have the content and just need LLM reasoning. - - ## Pre-loaded Documents Variable - - If documents were pre-loaded for this session, a `documents` variable is available: - ```python - # documents is a list of dicts with keys: id, title, uri, content - for doc in documents: - print(doc['title'], len(doc['content'])) - ``` - Check if it exists with: `if 'documents' in dir(): ...` - - ## Standard Library Modules - You can import any Python standard library module. - - ## Strategy Guide - - 1. **Explore First**: Start by listing documents or searching to understand what's available. Document names may differ from filenames (e.g., "tbmed593.pdf" might be stored as "TB MED 593" or similar). - 2. **If get_document returns None**: Use `list_documents()` to see actual document titles, or `search()` to find relevant content. - 3. **Iterative Refinement**: Run code, examine results, adjust your approach based on what you find. - 4. **Use print() Liberally**: The REPL captures stdout - print intermediate results to see what you're working with. - 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and use collections. - 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use llm(). - 7. **Cite Your Sources**: Track which documents/chunks informed your answer for citation. - - ## DoclingDocument API - - When you call `get_docling_document(id_or_title)`, you get a DoclingDocument object for structured document analysis. - - ### Properties - - `doc.texts` - List of all text items (paragraphs, headings, etc.) - - `doc.tables` - List of all tables - - `doc.pictures` - List of all pictures/figures - - `doc.name` - Document name - - ### Methods - - `doc.iterate_items(with_groups=False)` - Iterate all items with hierarchy level - Returns tuples of (item, level) where level is nesting depth - - `doc.export_to_markdown()` - Export entire document as markdown string - - ### Text Item Properties - - `item.text` - The text content - - `item.label` - Type: title, paragraph, section_header, list_item, etc. (lowercase enum values) - - `item.prov` - Provenance (page numbers, bounding boxes) - - ### Table Access - - `table.data.num_rows`, `table.data.num_cols` - Dimensions - - `table.data.table_cells` - List of TableCell objects - - `cell.text`, `cell.start_row_offset_idx`, `cell.start_col_offset_idx` - - ### Example Usage - ```python - doc = get_docling_document("My Document") - - # Get all headings - headings = [t.text for t in doc.texts if "header" in str(t.label)] - - # Iterate with structure - for item, level in doc.iterate_items(): - print(" " * level + item.text[:50]) - - # Extract table data - for table in doc.tables: - for cell in table.data.table_cells: - print(f"Row {cell.start_row_offset_idx}, Col {cell.start_col_offset_idx}: {cell.text}") - ``` - - ## Example Patterns - - ### Counting documents matching a condition - ```python - docs = list_documents(limit=100) - count = 0 - for doc in docs: - content = get_document(doc['id']) - if content and 'keyword' in content.lower(): - count += 1 - print(f"Found in: {doc['title']}") - print(f"Total: {count}") - ``` - - ### Aggregating data across documents - ```python - import re - numbers = [] - results = search("financial data", limit=20) - for r in results: - matches = re.findall(r'\$([\d,]+)', r['content']) - for m in matches: - numbers.append(int(m.replace(',', ''))) - print(f"Average: ${sum(numbers)/len(numbers):,.2f}") - ``` - - ### Using llm() for classification - ```python - # Get document content - content = get_document("Q1 Report") - # Use llm() to classify sentiment - sentiment = llm(f"Classify the sentiment as positive, negative, or mixed: {content}") - print(sentiment) - ``` - - ## Workflow - - 1. **ALWAYS start by using execute_code** to explore the knowledge base - 2. Run multiple code blocks as needed to gather information - 3. After collecting data, provide your final answer - - ## Output Format - - CRITICAL: Your final response MUST be valid JSON matching this exact schema: - ```json - {"answer": "Your complete answer here as a string", "program": "Your final consolidated program here as a string"} - ``` - - - `answer`: A clear answer to the user's question with key findings and references to specific documents/chunks. - - `program`: A single, self-contained Python program that produces the answer. Consolidate your exploratory code executions into one clean script. - - Do NOT return arbitrary JSON structures. Always use the exact format: {"answer": "...", "program": "..."} - - CRITICAL: You MUST call execute_code at least once before providing your answer. Never give up without trying to execute code first. - role: system - - content: What is the total revenue across all quarterly reports? - role: user - - content: |- - - Need to find quarterly reports, extract revenue amounts, sum. Likely documents titled like "Q1 Report", "Q2 Report"... Let's search "quarterly report revenue". - - role: assistant - tool_calls: - - function: - arguments: '{"code":"results = search(\"quarterly report revenue\", limit=20)\nprint(len(results))\nfor r in results[:5]:\n print(r[''document_title''], - r[''page_numbers''], r[''score''])\n"}' - name: execute_code - id: call_arzz3ioj - type: function - - content: '{"code":"results = search(\"quarterly report revenue\", limit=20)\nprint(len(results))\nfor r in results[:5]:\n print(r[''document_title''], - r[''page_numbers''], r[''score''])\n","stdout":"3\nQ1 Report [] 0.03226645663380623\nQ2 Report [] 0.03226645663380623\nQ3 - Report [] 0.032258063554763794\n","stderr":"","success":true}' - role: tool - tool_call_id: call_arzz3ioj - - content: |- - - We have 3 reports. Need to extract revenue. Let's fetch each document content and find revenue. Use regex for "$...". - - role: assistant - tool_calls: - - function: - arguments: '{"code":"import re\nreports = [''Q1 Report'', ''Q2 Report'', ''Q3 Report'']\nrevenues = {}\nfor rep - in reports:\n content = get_document(rep)\n if content is None:\n # try by title\n for doc - in list_documents():\n if rep.lower() in doc[''title''].lower():\n content = get_document(doc[''id''])\n break\n if - content:\n # find patterns like \"Revenue: $1,234,567\"\n matches = re.findall(r''Revenue[:\\s]*\\$([\\d,]+)'', - content)\n amounts = [int(m.replace('','','''')) for m in matches]\n if amounts:\n revenues[rep] - = sum(amounts)\n else:\n revenues[rep] = None\n else:\n revenues[rep] = None\nprint(revenues)\n"}' - name: execute_code - id: call_vuqzymvs - type: function - - content: '{"code":"import re\nreports = [''Q1 Report'', ''Q2 Report'', ''Q3 Report'']\nrevenues = {}\nfor rep in reports:\n content - = get_document(rep)\n if content is None:\n # try by title\n for doc in list_documents():\n if - rep.lower() in doc[''title''].lower():\n content = get_document(doc[''id''])\n break\n if - content:\n # find patterns like \"Revenue: $1,234,567\"\n matches = re.findall(r''Revenue[:\\s]*\\$([\\d,]+)'', - content)\n amounts = [int(m.replace('','','''')) for m in matches]\n if amounts:\n revenues[rep] - = sum(amounts)\n else:\n revenues[rep] = None\n else:\n revenues[rep] = None\nprint(revenues)\n","stdout":"{''Q1 - Report'': None, ''Q2 Report'': None, ''Q3 Report'': None}\n","stderr":"","success":true}' - role: tool - tool_call_id: call_vuqzymvs - - content: |2- - - No content or revenue not matched. Let's try list documents to see actual names and perhaps extract via pages. - role: assistant - tool_calls: - - function: - arguments: '{"code":"docs = list_documents(limit=50)\nfor d in docs:\n print(d[''title''])\n"}' - name: execute_code - id: call_z0mgsixf - type: function - - content: '{"code":"docs = list_documents(limit=50)\nfor d in docs:\n print(d[''title''])\n","stdout":"Q1 Report\nQ2 - Report\nQ3 Report\n","stderr":"","success":true}' - role: tool - tool_call_id: call_z0mgsixf - - content: |- - - Maybe revenue is in tables. Use docling. - - role: assistant - tool_calls: - - function: - arguments: '{"code":"doc = get_docling_document(''Q1 Report'')\nprint(doc.name)\n# print headings\nheadings = - [t.text for t in doc.texts if ''header'' in str(t.label)]\nprint(''headings:'', headlines:=headings[:5])\n# - find tables\nprint(f\"Tables: {len(doc.tables)}\")\nfor i,table in enumerate(doc.tables):\n print(f\"Table - {i} rows {table.data.num_rows} cols {table.data.num_cols}\")\n"}' - name: execute_code - id: call_acbn2mdz - type: function - - content: '{"code":"doc = get_docling_document(''Q1 Report'')\nprint(doc.name)\n# print headings\nheadings = [t.text - for t in doc.texts if ''header'' in str(t.label)]\nprint(''headings:'', headlines:=headings[:5])\n# find tables\nprint(f\"Tables: - {len(doc.tables)}\")\nfor i,table in enumerate(doc.tables):\n print(f\"Table {i} rows {table.data.num_rows} cols - {table.data.num_cols}\")\n","stdout":"content\nheadings: []\nTables: 0\n","stderr":"","success":true}' - role: tool - tool_call_id: call_acbn2mdz - - content: |- - - No tables. Probably revenue mentioned in paragraphs. Let's fetch entire content and search for "$". - - role: assistant - tool_calls: - - function: - arguments: '{"code":"content = get_document(''Q1 Report'')\nprint(content[:500])\nprint(''...'')\nprint(content.count(''$''))\n"}' - name: execute_code - id: call_y1fiwuj8 - type: function - - content: '{"code":"content = get_document(''Q1 Report'')\nprint(content[:500])\nprint(''...'')\nprint(content.count(''$''))\n","stdout":"Sales - report Q1: Revenue was $100,000.\n...\n1\n","stderr":"","success":true}' - role: tool - tool_call_id: call_y1fiwuj8 - model: gpt-oss - reasoning_effort: low - stream: false - tool_choice: auto - tools: - - function: - description: |- - Execute Python code in a Docker-sandboxed environment. - - The code has access to haiku.rag functions (search, list_documents, - get_document, get_docling_document, llm) and any Python standard - library module. - - Use print() to output results. - - Structured result with success status, stdout, and stderr. - - name: execute_code - parameters: - additionalProperties: false - properties: - code: - description: Python code to execute. - type: string - required: - - code - type: object - strict: true - type: function - - function: - description: Result from RLM agent execution. - name: final_result - parameters: - additionalProperties: false - properties: - answer: - description: The answer to the user's question - type: string - program: - description: The final consolidated program - type: string - required: - - answer - - program - type: object - strict: true - type: function - uri: http://localhost:11434/v1/chat/completions - response: - headers: - content-length: - - '646' - content-type: - - application/json - parsed_body: - choices: - - finish_reason: tool_calls - index: 0 - message: - content: '' - reasoning: Now extract $ amounts. - role: assistant - tool_calls: - - function: - arguments: '{"code":"import re\ncontent = get_document(''Q1 Report'')\namounts = re.findall(r''\\$(\\d{1,3}(?:,\\d{3})*)'', - content)\nprint(amounts)\nprint([int(a.replace('','','''')) for a in amounts])\n"}' - name: execute_code - id: call_i5cijc97 - index: 0 - type: function - created: 1770373360 - id: chatcmpl-728 - model: gpt-oss - object: chat.completion - system_fingerprint: fp_ollama - usage: - completion_tokens: 90 - prompt_tokens: 3045 - total_tokens: 3135 - status: - code: 200 - message: OK -- request: - headers: - accept: - - application/json - accept-encoding: - - gzip, deflate, zstd - connection: - - keep-alive - content-length: - - '13771' - content-type: - - application/json - host: - - localhost:11434 - method: POST - parsed_body: - messages: - - content: |- - You are a Recursive Language Model (RLM) agent that solves complex research questions by writing and executing Python code. - - IMPORTANT: You MUST use the `execute_code` tool to run Python code. The functions described below are ONLY available inside the execute_code tool - you cannot access them any other way. Always execute code to answer questions; do not just describe what code would do. - - CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just use them directly: - - search("query") ✓ CORRECT - - from haiku.rag import search ✗ WRONG - will fail - - You have access to a sandboxed Python environment with these haiku.rag functions (use them directly, no imports needed): - - ## Available Functions - - ### search(query, limit=10) -> list[dict] - Search the knowledge base using hybrid search (vector + full-text). - Returns list of dicts with keys: chunk_id, content, document_id, document_title, document_uri, score, page_numbers, headings - - ### list_documents(limit=10, offset=0) -> list[dict] - List available documents in the knowledge base. - Returns list of dicts with keys: id, title, uri, created_at - - ### get_document(id_or_title) -> str | None - Get the full text content of a document by ID, title, or URI. - Returns the document content as a string, or None if not found. - - ### get_docling_document(id_or_title) -> DoclingDocument | None - Get the structured DoclingDocument object for advanced analysis. - Returns a DoclingDocument object, or None if not found. - See "DoclingDocument API" section below for how to use it. - - ### llm(prompt) -> str - Call an LLM directly with the given prompt. Returns the response as a string. - Use this for classification, summarization, extraction, or any task where you - already have the content and just need LLM reasoning. - - ## Pre-loaded Documents Variable - - If documents were pre-loaded for this session, a `documents` variable is available: - ```python - # documents is a list of dicts with keys: id, title, uri, content - for doc in documents: - print(doc['title'], len(doc['content'])) - ``` - Check if it exists with: `if 'documents' in dir(): ...` - - ## Standard Library Modules - You can import any Python standard library module. - - ## Strategy Guide - - 1. **Explore First**: Start by listing documents or searching to understand what's available. Document names may differ from filenames (e.g., "tbmed593.pdf" might be stored as "TB MED 593" or similar). - 2. **If get_document returns None**: Use `list_documents()` to see actual document titles, or `search()` to find relevant content. - 3. **Iterative Refinement**: Run code, examine results, adjust your approach based on what you find. - 4. **Use print() Liberally**: The REPL captures stdout - print intermediate results to see what you're working with. - 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and use collections. - 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use llm(). - 7. **Cite Your Sources**: Track which documents/chunks informed your answer for citation. - - ## DoclingDocument API - - When you call `get_docling_document(id_or_title)`, you get a DoclingDocument object for structured document analysis. - - ### Properties - - `doc.texts` - List of all text items (paragraphs, headings, etc.) - - `doc.tables` - List of all tables - - `doc.pictures` - List of all pictures/figures - - `doc.name` - Document name - - ### Methods - - `doc.iterate_items(with_groups=False)` - Iterate all items with hierarchy level - Returns tuples of (item, level) where level is nesting depth - - `doc.export_to_markdown()` - Export entire document as markdown string - - ### Text Item Properties - - `item.text` - The text content - - `item.label` - Type: title, paragraph, section_header, list_item, etc. (lowercase enum values) - - `item.prov` - Provenance (page numbers, bounding boxes) - - ### Table Access - - `table.data.num_rows`, `table.data.num_cols` - Dimensions - - `table.data.table_cells` - List of TableCell objects - - `cell.text`, `cell.start_row_offset_idx`, `cell.start_col_offset_idx` - - ### Example Usage - ```python - doc = get_docling_document("My Document") - - # Get all headings - headings = [t.text for t in doc.texts if "header" in str(t.label)] - - # Iterate with structure - for item, level in doc.iterate_items(): - print(" " * level + item.text[:50]) - - # Extract table data - for table in doc.tables: - for cell in table.data.table_cells: - print(f"Row {cell.start_row_offset_idx}, Col {cell.start_col_offset_idx}: {cell.text}") - ``` - - ## Example Patterns - - ### Counting documents matching a condition - ```python - docs = list_documents(limit=100) - count = 0 - for doc in docs: - content = get_document(doc['id']) - if content and 'keyword' in content.lower(): - count += 1 - print(f"Found in: {doc['title']}") - print(f"Total: {count}") - ``` - - ### Aggregating data across documents - ```python - import re - numbers = [] - results = search("financial data", limit=20) - for r in results: - matches = re.findall(r'\$([\d,]+)', r['content']) - for m in matches: - numbers.append(int(m.replace(',', ''))) - print(f"Average: ${sum(numbers)/len(numbers):,.2f}") - ``` - - ### Using llm() for classification - ```python - # Get document content - content = get_document("Q1 Report") - # Use llm() to classify sentiment - sentiment = llm(f"Classify the sentiment as positive, negative, or mixed: {content}") - print(sentiment) - ``` - - ## Workflow - - 1. **ALWAYS start by using execute_code** to explore the knowledge base - 2. Run multiple code blocks as needed to gather information - 3. After collecting data, provide your final answer - - ## Output Format - - CRITICAL: Your final response MUST be valid JSON matching this exact schema: - ```json - {"answer": "Your complete answer here as a string", "program": "Your final consolidated program here as a string"} - ``` - - - `answer`: A clear answer to the user's question with key findings and references to specific documents/chunks. - - `program`: A single, self-contained Python program that produces the answer. Consolidate your exploratory code executions into one clean script. - - Do NOT return arbitrary JSON structures. Always use the exact format: {"answer": "...", "program": "..."} - - CRITICAL: You MUST call execute_code at least once before providing your answer. Never give up without trying to execute code first. - role: system - - content: What is the total revenue across all quarterly reports? - role: user - - content: |- - - Need to find quarterly reports, extract revenue amounts, sum. Likely documents titled like "Q1 Report", "Q2 Report"... Let's search "quarterly report revenue". - - role: assistant - tool_calls: - - function: - arguments: '{"code":"results = search(\"quarterly report revenue\", limit=20)\nprint(len(results))\nfor r in results[:5]:\n print(r[''document_title''], - r[''page_numbers''], r[''score''])\n"}' - name: execute_code - id: call_arzz3ioj - type: function - - content: '{"code":"results = search(\"quarterly report revenue\", limit=20)\nprint(len(results))\nfor r in results[:5]:\n print(r[''document_title''], - r[''page_numbers''], r[''score''])\n","stdout":"3\nQ1 Report [] 0.03226645663380623\nQ2 Report [] 0.03226645663380623\nQ3 - Report [] 0.032258063554763794\n","stderr":"","success":true}' - role: tool - tool_call_id: call_arzz3ioj - - content: |- - - We have 3 reports. Need to extract revenue. Let's fetch each document content and find revenue. Use regex for "$...". - - role: assistant - tool_calls: - - function: - arguments: '{"code":"import re\nreports = [''Q1 Report'', ''Q2 Report'', ''Q3 Report'']\nrevenues = {}\nfor rep - in reports:\n content = get_document(rep)\n if content is None:\n # try by title\n for doc - in list_documents():\n if rep.lower() in doc[''title''].lower():\n content = get_document(doc[''id''])\n break\n if - content:\n # find patterns like \"Revenue: $1,234,567\"\n matches = re.findall(r''Revenue[:\\s]*\\$([\\d,]+)'', - content)\n amounts = [int(m.replace('','','''')) for m in matches]\n if amounts:\n revenues[rep] - = sum(amounts)\n else:\n revenues[rep] = None\n else:\n revenues[rep] = None\nprint(revenues)\n"}' - name: execute_code - id: call_vuqzymvs - type: function - - content: '{"code":"import re\nreports = [''Q1 Report'', ''Q2 Report'', ''Q3 Report'']\nrevenues = {}\nfor rep in reports:\n content - = get_document(rep)\n if content is None:\n # try by title\n for doc in list_documents():\n if - rep.lower() in doc[''title''].lower():\n content = get_document(doc[''id''])\n break\n if - content:\n # find patterns like \"Revenue: $1,234,567\"\n matches = re.findall(r''Revenue[:\\s]*\\$([\\d,]+)'', - content)\n amounts = [int(m.replace('','','''')) for m in matches]\n if amounts:\n revenues[rep] - = sum(amounts)\n else:\n revenues[rep] = None\n else:\n revenues[rep] = None\nprint(revenues)\n","stdout":"{''Q1 - Report'': None, ''Q2 Report'': None, ''Q3 Report'': None}\n","stderr":"","success":true}' - role: tool - tool_call_id: call_vuqzymvs - - content: |2- - - No content or revenue not matched. Let's try list documents to see actual names and perhaps extract via pages. - role: assistant - tool_calls: - - function: - arguments: '{"code":"docs = list_documents(limit=50)\nfor d in docs:\n print(d[''title''])\n"}' - name: execute_code - id: call_z0mgsixf - type: function - - content: '{"code":"docs = list_documents(limit=50)\nfor d in docs:\n print(d[''title''])\n","stdout":"Q1 Report\nQ2 - Report\nQ3 Report\n","stderr":"","success":true}' - role: tool - tool_call_id: call_z0mgsixf - - content: |- - - Maybe revenue is in tables. Use docling. - - role: assistant - tool_calls: - - function: - arguments: '{"code":"doc = get_docling_document(''Q1 Report'')\nprint(doc.name)\n# print headings\nheadings = - [t.text for t in doc.texts if ''header'' in str(t.label)]\nprint(''headings:'', headlines:=headings[:5])\n# - find tables\nprint(f\"Tables: {len(doc.tables)}\")\nfor i,table in enumerate(doc.tables):\n print(f\"Table - {i} rows {table.data.num_rows} cols {table.data.num_cols}\")\n"}' - name: execute_code - id: call_acbn2mdz - type: function - - content: '{"code":"doc = get_docling_document(''Q1 Report'')\nprint(doc.name)\n# print headings\nheadings = [t.text - for t in doc.texts if ''header'' in str(t.label)]\nprint(''headings:'', headlines:=headings[:5])\n# find tables\nprint(f\"Tables: - {len(doc.tables)}\")\nfor i,table in enumerate(doc.tables):\n print(f\"Table {i} rows {table.data.num_rows} cols - {table.data.num_cols}\")\n","stdout":"content\nheadings: []\nTables: 0\n","stderr":"","success":true}' - role: tool - tool_call_id: call_acbn2mdz - - content: |- - - No tables. Probably revenue mentioned in paragraphs. Let's fetch entire content and search for "$". - - role: assistant - tool_calls: - - function: - arguments: '{"code":"content = get_document(''Q1 Report'')\nprint(content[:500])\nprint(''...'')\nprint(content.count(''$''))\n"}' - name: execute_code - id: call_y1fiwuj8 - type: function - - content: '{"code":"content = get_document(''Q1 Report'')\nprint(content[:500])\nprint(''...'')\nprint(content.count(''$''))\n","stdout":"Sales - report Q1: Revenue was $100,000.\n...\n1\n","stderr":"","success":true}' - role: tool - tool_call_id: call_y1fiwuj8 - - content: |- - - Now extract $ amounts. - - role: assistant - tool_calls: - - function: - arguments: '{"code":"import re\ncontent = get_document(''Q1 Report'')\namounts = re.findall(r''\\$(\\d{1,3}(?:,\\d{3})*)'', - content)\nprint(amounts)\nprint([int(a.replace('','','''')) for a in amounts])\n"}' - name: execute_code - id: call_i5cijc97 - type: function - - content: '{"code":"import re\ncontent = get_document(''Q1 Report'')\namounts = re.findall(r''\\$(\\d{1,3}(?:,\\d{3})*)'', - content)\nprint(amounts)\nprint([int(a.replace('','','''')) for a in amounts])\n","stdout":"[''100,000'']\n[100000]\n","stderr":"","success":true}' - role: tool - tool_call_id: call_i5cijc97 - model: gpt-oss - reasoning_effort: low - stream: false - tool_choice: auto - tools: - - function: - description: |- - Execute Python code in a Docker-sandboxed environment. - - The code has access to haiku.rag functions (search, list_documents, - get_document, get_docling_document, llm) and any Python standard - library module. - - Use print() to output results. - - Structured result with success status, stdout, and stderr. - - name: execute_code - parameters: - additionalProperties: false - properties: - code: - description: Python code to execute. - type: string - required: - - code - type: object - strict: true - type: function - - function: - description: Result from RLM agent execution. - name: final_result - parameters: - additionalProperties: false - properties: - answer: - description: The answer to the user's question - type: string - program: - description: The final consolidated program - type: string - required: - - answer - - program - type: object - strict: true - type: function - uri: http://localhost:11434/v1/chat/completions - response: - headers: - content-length: - - '875' - content-type: - - application/json - parsed_body: - choices: - - finish_reason: tool_calls - index: 0 - message: - content: '' - reasoning: Do same for Q2 and Q3. - role: assistant - tool_calls: - - function: - arguments: '{"code":"import re\nreports = [''Q1 Report'',''Q2 Report'',''Q3 Report'']\nvalues={}\nfor r in reports:\n c=get_document(r)\n if - c:\n amounts=re.findall(r''\\$(\\d{1,3}(?:,\\d{3})*)'',c)\n totals=sum(int(a.replace('','','''')) - for a in amounts)\n values[r]=totals\n else:\n values[r]=None\nprint(values)\nprint(''Total - revenue:'',sum(v for v in values.values() if v is not None))\n"}' - name: execute_code - id: call_4f61ktvo - index: 0 - type: function - created: 1770373364 - id: chatcmpl-865 - model: gpt-oss - object: chat.completion - system_fingerprint: fp_ollama - usage: - completion_tokens: 158 - prompt_tokens: 3239 - total_tokens: 3397 - status: - code: 200 - message: OK -- request: - headers: - accept: - - application/json - accept-encoding: - - gzip, deflate, zstd - connection: - - keep-alive - content-length: - - '15009' - content-type: - - application/json - host: - - localhost:11434 - method: POST - parsed_body: - messages: - - content: |- - You are a Recursive Language Model (RLM) agent that solves complex research questions by writing and executing Python code. - - IMPORTANT: You MUST use the `execute_code` tool to run Python code. The functions described below are ONLY available inside the execute_code tool - you cannot access them any other way. Always execute code to answer questions; do not just describe what code would do. - - CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just use them directly: - - search("query") ✓ CORRECT - - from haiku.rag import search ✗ WRONG - will fail - - You have access to a sandboxed Python environment with these haiku.rag functions (use them directly, no imports needed): - - ## Available Functions - - ### search(query, limit=10) -> list[dict] - Search the knowledge base using hybrid search (vector + full-text). - Returns list of dicts with keys: chunk_id, content, document_id, document_title, document_uri, score, page_numbers, headings - - ### list_documents(limit=10, offset=0) -> list[dict] - List available documents in the knowledge base. - Returns list of dicts with keys: id, title, uri, created_at - - ### get_document(id_or_title) -> str | None - Get the full text content of a document by ID, title, or URI. - Returns the document content as a string, or None if not found. - - ### get_docling_document(id_or_title) -> DoclingDocument | None - Get the structured DoclingDocument object for advanced analysis. - Returns a DoclingDocument object, or None if not found. - See "DoclingDocument API" section below for how to use it. - - ### llm(prompt) -> str - Call an LLM directly with the given prompt. Returns the response as a string. - Use this for classification, summarization, extraction, or any task where you - already have the content and just need LLM reasoning. - - ## Pre-loaded Documents Variable - - If documents were pre-loaded for this session, a `documents` variable is available: - ```python - # documents is a list of dicts with keys: id, title, uri, content - for doc in documents: - print(doc['title'], len(doc['content'])) - ``` - Check if it exists with: `if 'documents' in dir(): ...` - - ## Standard Library Modules - You can import any Python standard library module. - - ## Strategy Guide - - 1. **Explore First**: Start by listing documents or searching to understand what's available. Document names may differ from filenames (e.g., "tbmed593.pdf" might be stored as "TB MED 593" or similar). - 2. **If get_document returns None**: Use `list_documents()` to see actual document titles, or `search()` to find relevant content. - 3. **Iterative Refinement**: Run code, examine results, adjust your approach based on what you find. - 4. **Use print() Liberally**: The REPL captures stdout - print intermediate results to see what you're working with. - 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and use collections. - 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use llm(). - 7. **Cite Your Sources**: Track which documents/chunks informed your answer for citation. - - ## DoclingDocument API - - When you call `get_docling_document(id_or_title)`, you get a DoclingDocument object for structured document analysis. - - ### Properties - - `doc.texts` - List of all text items (paragraphs, headings, etc.) - - `doc.tables` - List of all tables - - `doc.pictures` - List of all pictures/figures - - `doc.name` - Document name - - ### Methods - - `doc.iterate_items(with_groups=False)` - Iterate all items with hierarchy level - Returns tuples of (item, level) where level is nesting depth - - `doc.export_to_markdown()` - Export entire document as markdown string - - ### Text Item Properties - - `item.text` - The text content - - `item.label` - Type: title, paragraph, section_header, list_item, etc. (lowercase enum values) - - `item.prov` - Provenance (page numbers, bounding boxes) - - ### Table Access - - `table.data.num_rows`, `table.data.num_cols` - Dimensions - - `table.data.table_cells` - List of TableCell objects - - `cell.text`, `cell.start_row_offset_idx`, `cell.start_col_offset_idx` - - ### Example Usage - ```python - doc = get_docling_document("My Document") - - # Get all headings - headings = [t.text for t in doc.texts if "header" in str(t.label)] - - # Iterate with structure - for item, level in doc.iterate_items(): - print(" " * level + item.text[:50]) - - # Extract table data - for table in doc.tables: - for cell in table.data.table_cells: - print(f"Row {cell.start_row_offset_idx}, Col {cell.start_col_offset_idx}: {cell.text}") - ``` - - ## Example Patterns - - ### Counting documents matching a condition - ```python - docs = list_documents(limit=100) - count = 0 - for doc in docs: - content = get_document(doc['id']) - if content and 'keyword' in content.lower(): - count += 1 - print(f"Found in: {doc['title']}") - print(f"Total: {count}") - ``` - - ### Aggregating data across documents - ```python - import re - numbers = [] - results = search("financial data", limit=20) - for r in results: - matches = re.findall(r'\$([\d,]+)', r['content']) - for m in matches: - numbers.append(int(m.replace(',', ''))) - print(f"Average: ${sum(numbers)/len(numbers):,.2f}") - ``` - - ### Using llm() for classification - ```python - # Get document content - content = get_document("Q1 Report") - # Use llm() to classify sentiment - sentiment = llm(f"Classify the sentiment as positive, negative, or mixed: {content}") - print(sentiment) - ``` - - ## Workflow - - 1. **ALWAYS start by using execute_code** to explore the knowledge base - 2. Run multiple code blocks as needed to gather information - 3. After collecting data, provide your final answer - - ## Output Format - - CRITICAL: Your final response MUST be valid JSON matching this exact schema: - ```json - {"answer": "Your complete answer here as a string", "program": "Your final consolidated program here as a string"} - ``` - - - `answer`: A clear answer to the user's question with key findings and references to specific documents/chunks. - - `program`: A single, self-contained Python program that produces the answer. Consolidate your exploratory code executions into one clean script. - - Do NOT return arbitrary JSON structures. Always use the exact format: {"answer": "...", "program": "..."} - - CRITICAL: You MUST call execute_code at least once before providing your answer. Never give up without trying to execute code first. - role: system - - content: What is the total revenue across all quarterly reports? - role: user - - content: |- - - Need to find quarterly reports, extract revenue amounts, sum. Likely documents titled like "Q1 Report", "Q2 Report"... Let's search "quarterly report revenue". - - role: assistant - tool_calls: - - function: - arguments: '{"code":"results = search(\"quarterly report revenue\", limit=20)\nprint(len(results))\nfor r in results[:5]:\n print(r[''document_title''], - r[''page_numbers''], r[''score''])\n"}' - name: execute_code - id: call_arzz3ioj - type: function - - content: '{"code":"results = search(\"quarterly report revenue\", limit=20)\nprint(len(results))\nfor r in results[:5]:\n print(r[''document_title''], - r[''page_numbers''], r[''score''])\n","stdout":"3\nQ1 Report [] 0.03226645663380623\nQ2 Report [] 0.03226645663380623\nQ3 - Report [] 0.032258063554763794\n","stderr":"","success":true}' - role: tool - tool_call_id: call_arzz3ioj - - content: |- - - We have 3 reports. Need to extract revenue. Let's fetch each document content and find revenue. Use regex for "$...". - - role: assistant - tool_calls: - - function: - arguments: '{"code":"import re\nreports = [''Q1 Report'', ''Q2 Report'', ''Q3 Report'']\nrevenues = {}\nfor rep - in reports:\n content = get_document(rep)\n if content is None:\n # try by title\n for doc - in list_documents():\n if rep.lower() in doc[''title''].lower():\n content = get_document(doc[''id''])\n break\n if - content:\n # find patterns like \"Revenue: $1,234,567\"\n matches = re.findall(r''Revenue[:\\s]*\\$([\\d,]+)'', - content)\n amounts = [int(m.replace('','','''')) for m in matches]\n if amounts:\n revenues[rep] - = sum(amounts)\n else:\n revenues[rep] = None\n else:\n revenues[rep] = None\nprint(revenues)\n"}' - name: execute_code - id: call_vuqzymvs - type: function - - content: '{"code":"import re\nreports = [''Q1 Report'', ''Q2 Report'', ''Q3 Report'']\nrevenues = {}\nfor rep in reports:\n content - = get_document(rep)\n if content is None:\n # try by title\n for doc in list_documents():\n if - rep.lower() in doc[''title''].lower():\n content = get_document(doc[''id''])\n break\n if - content:\n # find patterns like \"Revenue: $1,234,567\"\n matches = re.findall(r''Revenue[:\\s]*\\$([\\d,]+)'', - content)\n amounts = [int(m.replace('','','''')) for m in matches]\n if amounts:\n revenues[rep] - = sum(amounts)\n else:\n revenues[rep] = None\n else:\n revenues[rep] = None\nprint(revenues)\n","stdout":"{''Q1 - Report'': None, ''Q2 Report'': None, ''Q3 Report'': None}\n","stderr":"","success":true}' - role: tool - tool_call_id: call_vuqzymvs - - content: |2- - - No content or revenue not matched. Let's try list documents to see actual names and perhaps extract via pages. - role: assistant - tool_calls: - - function: - arguments: '{"code":"docs = list_documents(limit=50)\nfor d in docs:\n print(d[''title''])\n"}' - name: execute_code - id: call_z0mgsixf - type: function - - content: '{"code":"docs = list_documents(limit=50)\nfor d in docs:\n print(d[''title''])\n","stdout":"Q1 Report\nQ2 - Report\nQ3 Report\n","stderr":"","success":true}' - role: tool - tool_call_id: call_z0mgsixf - - content: |- - - Maybe revenue is in tables. Use docling. - - role: assistant - tool_calls: - - function: - arguments: '{"code":"doc = get_docling_document(''Q1 Report'')\nprint(doc.name)\n# print headings\nheadings = - [t.text for t in doc.texts if ''header'' in str(t.label)]\nprint(''headings:'', headlines:=headings[:5])\n# - find tables\nprint(f\"Tables: {len(doc.tables)}\")\nfor i,table in enumerate(doc.tables):\n print(f\"Table - {i} rows {table.data.num_rows} cols {table.data.num_cols}\")\n"}' - name: execute_code - id: call_acbn2mdz - type: function - - content: '{"code":"doc = get_docling_document(''Q1 Report'')\nprint(doc.name)\n# print headings\nheadings = [t.text - for t in doc.texts if ''header'' in str(t.label)]\nprint(''headings:'', headlines:=headings[:5])\n# find tables\nprint(f\"Tables: - {len(doc.tables)}\")\nfor i,table in enumerate(doc.tables):\n print(f\"Table {i} rows {table.data.num_rows} cols - {table.data.num_cols}\")\n","stdout":"content\nheadings: []\nTables: 0\n","stderr":"","success":true}' - role: tool - tool_call_id: call_acbn2mdz - - content: |- - - No tables. Probably revenue mentioned in paragraphs. Let's fetch entire content and search for "$". - - role: assistant - tool_calls: - - function: - arguments: '{"code":"content = get_document(''Q1 Report'')\nprint(content[:500])\nprint(''...'')\nprint(content.count(''$''))\n"}' - name: execute_code - id: call_y1fiwuj8 - type: function - - content: '{"code":"content = get_document(''Q1 Report'')\nprint(content[:500])\nprint(''...'')\nprint(content.count(''$''))\n","stdout":"Sales - report Q1: Revenue was $100,000.\n...\n1\n","stderr":"","success":true}' - role: tool - tool_call_id: call_y1fiwuj8 - - content: |- - - Now extract $ amounts. - - role: assistant - tool_calls: - - function: - arguments: '{"code":"import re\ncontent = get_document(''Q1 Report'')\namounts = re.findall(r''\\$(\\d{1,3}(?:,\\d{3})*)'', - content)\nprint(amounts)\nprint([int(a.replace('','','''')) for a in amounts])\n"}' - name: execute_code - id: call_i5cijc97 - type: function - - content: '{"code":"import re\ncontent = get_document(''Q1 Report'')\namounts = re.findall(r''\\$(\\d{1,3}(?:,\\d{3})*)'', - content)\nprint(amounts)\nprint([int(a.replace('','','''')) for a in amounts])\n","stdout":"[''100,000'']\n[100000]\n","stderr":"","success":true}' - role: tool - tool_call_id: call_i5cijc97 - - content: |- - - Do same for Q2 and Q3. - - role: assistant - tool_calls: - - function: - arguments: '{"code":"import re\nreports = [''Q1 Report'',''Q2 Report'',''Q3 Report'']\nvalues={}\nfor r in reports:\n c=get_document(r)\n if - c:\n amounts=re.findall(r''\\$(\\d{1,3}(?:,\\d{3})*)'',c)\n totals=sum(int(a.replace('','','''')) - for a in amounts)\n values[r]=totals\n else:\n values[r]=None\nprint(values)\nprint(''Total - revenue:'',sum(v for v in values.values() if v is not None))\n"}' - name: execute_code - id: call_4f61ktvo - type: function - - content: '{"code":"import re\nreports = [''Q1 Report'',''Q2 Report'',''Q3 Report'']\nvalues={}\nfor r in reports:\n c=get_document(r)\n if - c:\n amounts=re.findall(r''\\$(\\d{1,3}(?:,\\d{3})*)'',c)\n totals=sum(int(a.replace('','','''')) - for a in amounts)\n values[r]=totals\n else:\n values[r]=None\nprint(values)\nprint(''Total revenue:'',sum(v - for v in values.values() if v is not None))\n","stdout":"{''Q1 Report'': 100000, ''Q2 Report'': 150000, ''Q3 Report'': - 200000}\nTotal revenue: 450000\n","stderr":"","success":true}' - role: tool - tool_call_id: call_4f61ktvo - model: gpt-oss - reasoning_effort: low - stream: false - tool_choice: auto - tools: - - function: - description: |- - Execute Python code in a Docker-sandboxed environment. - - The code has access to haiku.rag functions (search, list_documents, - get_document, get_docling_document, llm) and any Python standard - library module. - - Use print() to output results. - - Structured result with success status, stdout, and stderr. - - name: execute_code - parameters: - additionalProperties: false - properties: - code: - description: Python code to execute. - type: string - required: - - code - type: object - strict: true - type: function - - function: - description: Result from RLM agent execution. - name: final_result - parameters: - additionalProperties: false - properties: - answer: - description: The answer to the user's question - type: string - program: - description: The final consolidated program - type: string - required: - - answer - - program - type: object - strict: true - type: function - uri: http://localhost:11434/v1/chat/completions - response: - headers: - content-length: - - '1016' + - '1421' content-type: - application/json parsed_body: @@ -2671,22 +912,27 @@ interactions: - finish_reason: stop index: 0 message: - content: '{"answer":"The combined revenue reported across the three quarterly reports equals $450,000. Individual - figures are:\n• Q1 Report: $100,000\n• Q2 Report: $150,000\n• Q3 Report: $200,000\n\nSources are the full text - of each report where the amounts are stated in the revenue sentences.","program":"import re\nreports = [''Q1 Report'',''Q2 - Report'',''Q3 Report'']\nvalues={}\nfor r in reports:\n c=get_document(r)\n if c:\n amounts=re.findall(r''\\$(\\d{1,3}(?:,\\d{3})*)'',c)\n totals=sum(int(a.replace('','','''')) - for a in amounts)\n values[r]=totals\n else:\n values[r]=None\nprint(values)\nprint(''Total revenue:'',sum(v - for v in values.values() if v is not None))"}' + content: '{"answer":"The summed revenue reported across all available quarterly reports (Q1, Q2, and Q3) is **$450,000**. + The figures are:\n- **Q1 Report** – $100,000 (citation: chunk fb347b1c‑9bf0‑455b‑8535‑3484f9578731)\n- **Q2 Report** + – $150,000 (citation: chunk 50f60a6f‑53e4‑4850‑8d5b‑a3c00d4375c6)\n- **Q3 Report** – $200,000 (citation: chunk + 70233263‑be1a‑414e‑8841‑c9509076530e)\n\nNo Q4 report was found in the knowledge base. The total is therefore + 100 k + 150 k + 200 k = 450 k dollars.", "program":"# Calculate total revenue from quarterly reports\nresults + = await search(\"sales report revenue\", limit=10)\nrevenue = 0\nfor r in results:\n # Use llm to extract dollar + amount\n amount_text = await llm(f\"Extract the dollar amount from this snippet: {r[''content'']}\")\n # + Clean and convert\n amt = int(''''.join([c for c in amount_text if c.isdigit()]))\n revenue += amt\nprint(f\"Total + revenue: ${revenue}\")"}' + reasoning: Results show Q1 100k, Q2 150k, Q3 200k. No Q4? Maybe not included. Total revenue = 450k. Provide citations + from chunks. role: assistant - created: 1770373369 - id: chatcmpl-835 + created: 1771924512 + id: chatcmpl-210 model: gpt-oss object: chat.completion system_fingerprint: fp_ollama usage: - completion_tokens: 206 - prompt_tokens: 3588 - total_tokens: 3794 + completion_tokens: 367 + prompt_tokens: 2311 + total_tokens: 2678 status: code: 200 message: OK diff --git a/tests/cassettes/test_rlm/TestClientRLMIntegration.test_rlm_count_documents.yaml b/tests/cassettes/test_rlm/TestClientRLMIntegration.test_rlm_count_documents.yaml index 7948eff4..ae352e94 100644 --- a/tests/cassettes/test_rlm/TestClientRLMIntegration.test_rlm_count_documents.yaml +++ b/tests/cassettes/test_rlm/TestClientRLMIntegration.test_rlm_count_documents.yaml @@ -128,7 +128,7 @@ interactions: connection: - keep-alive content-length: - - '7774' + - '7325' content-type: - application/json host: @@ -141,32 +141,33 @@ interactions: IMPORTANT: You MUST use the `execute_code` tool to run Python code. The functions described below are ONLY available inside the execute_code tool - you cannot access them any other way. Always execute code to answer questions; do not just describe what code would do. - CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just use them directly: - - search("query") ✓ CORRECT + CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just call them with `await`: + - results = await search("query") ✓ CORRECT - from haiku.rag import search ✗ WRONG - will fail + - results = search("query") ✗ WRONG - must use await - You have access to a sandboxed Python environment with these haiku.rag functions (use them directly, no imports needed): + You have access to a sandboxed Python interpreter with these haiku.rag functions (use them directly with `await`, no imports needed): ## Available Functions - ### search(query, limit=10) -> list[dict] + ### await search(query, limit=10) -> list[dict] Search the knowledge base using hybrid search (vector + full-text). Returns list of dicts with keys: chunk_id, content, document_id, document_title, document_uri, score, page_numbers, headings - ### list_documents(limit=10, offset=0) -> list[dict] + ### await list_documents(limit=10, offset=0) -> list[dict] List available documents in the knowledge base. Returns list of dicts with keys: id, title, uri, created_at - ### get_document(id_or_title) -> str | None + ### await get_document(id_or_title) -> str | None Get the full text content of a document by ID, title, or URI. Returns the document content as a string, or None if not found. - ### get_docling_document(id_or_title) -> DoclingDocument | None - Get the structured DoclingDocument object for advanced analysis. - Returns a DoclingDocument object, or None if not found. - See "DoclingDocument API" section below for how to use it. + ### await get_chunk(chunk_id) -> dict | None + Get a specific chunk by its ID (from search results). + Returns dict with keys: chunk_id, content, document_id, document_title, headings, page_numbers, labels + Use this to retrieve full chunk details and metadata for citation. - ### llm(prompt) -> str + ### await llm(prompt) -> str Call an LLM directly with the given prompt. Returns the response as a string. Use this for classification, summarization, extraction, or any task where you already have the content and just need LLM reasoning. @@ -181,93 +182,64 @@ interactions: ``` Check if it exists with: `if 'documents' in dir(): ...` - ## Standard Library Modules - You can import any Python standard library module. + ## Available Python Features + + The interpreter supports: variables, arithmetic, strings, f-strings, lists, dicts, tuples, sets, loops, conditionals, comprehensions, functions, async/await, `map()`, `sorted()`/`.sort(key=...)`, try/except, and the `json` module. + + Not supported: imports (other than `json`), class definitions, generators/yield, match statements, decorators, `with` statements. + + For pattern matching or text extraction, use string methods (`str.split`, `str.find`, `str.startswith`, `in` operator) or the `llm()` function. ## Strategy Guide 1. **Explore First**: Start by listing documents or searching to understand what's available. Document names may differ from filenames (e.g., "tbmed593.pdf" might be stored as "TB MED 593" or similar). - 2. **If get_document returns None**: Use `list_documents()` to see actual document titles, or `search()` to find relevant content. + 2. **If get_document returns None**: Use `await list_documents()` to see actual document titles, or `await search()` to find relevant content. 3. **Iterative Refinement**: Run code, examine results, adjust your approach based on what you find. - 4. **Use print() Liberally**: The REPL captures stdout - print intermediate results to see what you're working with. - 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and use collections. - 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use llm(). - 7. **Cite Your Sources**: Track which documents/chunks informed your answer for citation. - - ## DoclingDocument API - - When you call `get_docling_document(id_or_title)`, you get a DoclingDocument object for structured document analysis. - - ### Properties - - `doc.texts` - List of all text items (paragraphs, headings, etc.) - - `doc.tables` - List of all tables - - `doc.pictures` - List of all pictures/figures - - `doc.name` - Document name - - ### Methods - - `doc.iterate_items(with_groups=False)` - Iterate all items with hierarchy level - Returns tuples of (item, level) where level is nesting depth - - `doc.export_to_markdown()` - Export entire document as markdown string - - ### Text Item Properties - - `item.text` - The text content - - `item.label` - Type: title, paragraph, section_header, list_item, etc. (lowercase enum values) - - `item.prov` - Provenance (page numbers, bounding boxes) - - ### Table Access - - `table.data.num_rows`, `table.data.num_cols` - Dimensions - - `table.data.table_cells` - List of TableCell objects - - `cell.text`, `cell.start_row_offset_idx`, `cell.start_col_offset_idx` - - ### Example Usage - ```python - doc = get_docling_document("My Document") - - # Get all headings - headings = [t.text for t in doc.texts if "header" in str(t.label)] - - # Iterate with structure - for item, level in doc.iterate_items(): - print(" " * level + item.text[:50]) - - # Extract table data - for table in doc.tables: - for cell in table.data.table_cells: - print(f"Row {cell.start_row_offset_idx}, Col {cell.start_col_offset_idx}: {cell.text}") - ``` + 4. **Use print() Liberally**: The sandbox captures stdout - print intermediate results to see what you're working with. + 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and data structures. + 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use `await llm()`. + 7. **Cite Your Sources**: Use get_chunk() to retrieve chunk metadata for citations. Track which documents/chunks informed your answer. ## Example Patterns ### Counting documents matching a condition ```python - docs = list_documents(limit=100) + docs = await list_documents(limit=100) count = 0 for doc in docs: - content = get_document(doc['id']) + content = await get_document(doc['id']) if content and 'keyword' in content.lower(): count += 1 print(f"Found in: {doc['title']}") print(f"Total: {count}") ``` - ### Aggregating data across documents + ### Extracting data with llm() ```python - import re numbers = [] - results = search("financial data", limit=20) + results = await search("financial data", limit=20) for r in results: - matches = re.findall(r'\$([\d,]+)', r['content']) - for m in matches: - numbers.append(int(m.replace(',', ''))) - print(f"Average: ${sum(numbers)/len(numbers):,.2f}") + extracted = await llm(f"Extract all dollar amounts from this text as a comma-separated list of numbers (no $ signs): {r['content']}") + for part in extracted.split(','): + part = part.strip().replace(',', '') + if part.isdigit(): + numbers.append(int(part)) + if numbers: + print(f"Average: {sum(numbers) / len(numbers)}") + ``` + + ### Using search results with get_chunk for citations + ```python + results = await search("safety requirements", limit=5) + for r in results: + chunk = await get_chunk(r['chunk_id']) + print(f"From '{chunk['document_title']}', page {chunk['page_numbers']}: {chunk['content'][:100]}") ``` ### Using llm() for classification ```python - # Get document content - content = get_document("Q1 Report") - # Use llm() to classify sentiment - sentiment = llm(f"Classify the sentiment as positive, negative, or mixed: {content}") + content = await get_document("Q1 Report") + sentiment = await llm(f"Classify the sentiment as positive, negative, or mixed: {content}") print(sentiment) ``` @@ -300,11 +272,10 @@ interactions: tools: - function: description: |- - Execute Python code in a Docker-sandboxed environment. + Execute Python code in a sandboxed interpreter. The code has access to haiku.rag functions (search, list_documents, - get_document, get_docling_document, llm) and any Python standard - library module. + get_document, get_chunk, llm). Use print() to output results. @@ -344,7 +315,7 @@ interactions: response: headers: content-length: - - '523' + - '514' content-type: - application/json parsed_body: @@ -353,24 +324,24 @@ interactions: index: 0 message: content: '' - reasoning: We need to list documents. + reasoning: Need to list docs. role: assistant tool_calls: - function: - arguments: '{"code":"# list documents\nimport json\nprint(list_documents())\n"}' + arguments: '{"code":"docs = await list_documents(limit=1000)\nprint(len(docs))"}' name: execute_code - id: call_d8xhmimu + id: call_stp0fimx index: 0 type: function - created: 1770373335 - id: chatcmpl-184 + created: 1771924497 + id: chatcmpl-750 model: gpt-oss object: chat.completion system_fingerprint: fp_ollama usage: - completion_tokens: 42 - prompt_tokens: 1747 - total_tokens: 1789 + completion_tokens: 44 + prompt_tokens: 1623 + total_tokens: 1667 status: code: 200 message: OK @@ -383,7 +354,7 @@ interactions: connection: - keep-alive content-length: - - '8588' + - '7759' content-type: - application/json host: @@ -396,32 +367,33 @@ interactions: IMPORTANT: You MUST use the `execute_code` tool to run Python code. The functions described below are ONLY available inside the execute_code tool - you cannot access them any other way. Always execute code to answer questions; do not just describe what code would do. - CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just use them directly: - - search("query") ✓ CORRECT + CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just call them with `await`: + - results = await search("query") ✓ CORRECT - from haiku.rag import search ✗ WRONG - will fail + - results = search("query") ✗ WRONG - must use await - You have access to a sandboxed Python environment with these haiku.rag functions (use them directly, no imports needed): + You have access to a sandboxed Python interpreter with these haiku.rag functions (use them directly with `await`, no imports needed): ## Available Functions - ### search(query, limit=10) -> list[dict] + ### await search(query, limit=10) -> list[dict] Search the knowledge base using hybrid search (vector + full-text). Returns list of dicts with keys: chunk_id, content, document_id, document_title, document_uri, score, page_numbers, headings - ### list_documents(limit=10, offset=0) -> list[dict] + ### await list_documents(limit=10, offset=0) -> list[dict] List available documents in the knowledge base. Returns list of dicts with keys: id, title, uri, created_at - ### get_document(id_or_title) -> str | None + ### await get_document(id_or_title) -> str | None Get the full text content of a document by ID, title, or URI. Returns the document content as a string, or None if not found. - ### get_docling_document(id_or_title) -> DoclingDocument | None - Get the structured DoclingDocument object for advanced analysis. - Returns a DoclingDocument object, or None if not found. - See "DoclingDocument API" section below for how to use it. + ### await get_chunk(chunk_id) -> dict | None + Get a specific chunk by its ID (from search results). + Returns dict with keys: chunk_id, content, document_id, document_title, headings, page_numbers, labels + Use this to retrieve full chunk details and metadata for citation. - ### llm(prompt) -> str + ### await llm(prompt) -> str Call an LLM directly with the given prompt. Returns the response as a string. Use this for classification, summarization, extraction, or any task where you already have the content and just need LLM reasoning. @@ -436,93 +408,64 @@ interactions: ``` Check if it exists with: `if 'documents' in dir(): ...` - ## Standard Library Modules - You can import any Python standard library module. + ## Available Python Features + + The interpreter supports: variables, arithmetic, strings, f-strings, lists, dicts, tuples, sets, loops, conditionals, comprehensions, functions, async/await, `map()`, `sorted()`/`.sort(key=...)`, try/except, and the `json` module. + + Not supported: imports (other than `json`), class definitions, generators/yield, match statements, decorators, `with` statements. + + For pattern matching or text extraction, use string methods (`str.split`, `str.find`, `str.startswith`, `in` operator) or the `llm()` function. ## Strategy Guide 1. **Explore First**: Start by listing documents or searching to understand what's available. Document names may differ from filenames (e.g., "tbmed593.pdf" might be stored as "TB MED 593" or similar). - 2. **If get_document returns None**: Use `list_documents()` to see actual document titles, or `search()` to find relevant content. + 2. **If get_document returns None**: Use `await list_documents()` to see actual document titles, or `await search()` to find relevant content. 3. **Iterative Refinement**: Run code, examine results, adjust your approach based on what you find. - 4. **Use print() Liberally**: The REPL captures stdout - print intermediate results to see what you're working with. - 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and use collections. - 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use llm(). - 7. **Cite Your Sources**: Track which documents/chunks informed your answer for citation. - - ## DoclingDocument API - - When you call `get_docling_document(id_or_title)`, you get a DoclingDocument object for structured document analysis. - - ### Properties - - `doc.texts` - List of all text items (paragraphs, headings, etc.) - - `doc.tables` - List of all tables - - `doc.pictures` - List of all pictures/figures - - `doc.name` - Document name - - ### Methods - - `doc.iterate_items(with_groups=False)` - Iterate all items with hierarchy level - Returns tuples of (item, level) where level is nesting depth - - `doc.export_to_markdown()` - Export entire document as markdown string - - ### Text Item Properties - - `item.text` - The text content - - `item.label` - Type: title, paragraph, section_header, list_item, etc. (lowercase enum values) - - `item.prov` - Provenance (page numbers, bounding boxes) - - ### Table Access - - `table.data.num_rows`, `table.data.num_cols` - Dimensions - - `table.data.table_cells` - List of TableCell objects - - `cell.text`, `cell.start_row_offset_idx`, `cell.start_col_offset_idx` - - ### Example Usage - ```python - doc = get_docling_document("My Document") - - # Get all headings - headings = [t.text for t in doc.texts if "header" in str(t.label)] - - # Iterate with structure - for item, level in doc.iterate_items(): - print(" " * level + item.text[:50]) - - # Extract table data - for table in doc.tables: - for cell in table.data.table_cells: - print(f"Row {cell.start_row_offset_idx}, Col {cell.start_col_offset_idx}: {cell.text}") - ``` + 4. **Use print() Liberally**: The sandbox captures stdout - print intermediate results to see what you're working with. + 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and data structures. + 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use `await llm()`. + 7. **Cite Your Sources**: Use get_chunk() to retrieve chunk metadata for citations. Track which documents/chunks informed your answer. ## Example Patterns ### Counting documents matching a condition ```python - docs = list_documents(limit=100) + docs = await list_documents(limit=100) count = 0 for doc in docs: - content = get_document(doc['id']) + content = await get_document(doc['id']) if content and 'keyword' in content.lower(): count += 1 print(f"Found in: {doc['title']}") print(f"Total: {count}") ``` - ### Aggregating data across documents + ### Extracting data with llm() ```python - import re numbers = [] - results = search("financial data", limit=20) + results = await search("financial data", limit=20) for r in results: - matches = re.findall(r'\$([\d,]+)', r['content']) - for m in matches: - numbers.append(int(m.replace(',', ''))) - print(f"Average: ${sum(numbers)/len(numbers):,.2f}") + extracted = await llm(f"Extract all dollar amounts from this text as a comma-separated list of numbers (no $ signs): {r['content']}") + for part in extracted.split(','): + part = part.strip().replace(',', '') + if part.isdigit(): + numbers.append(int(part)) + if numbers: + print(f"Average: {sum(numbers) / len(numbers)}") + ``` + + ### Using search results with get_chunk for citations + ```python + results = await search("safety requirements", limit=5) + for r in results: + chunk = await get_chunk(r['chunk_id']) + print(f"From '{chunk['document_title']}', page {chunk['page_numbers']}: {chunk['content'][:100]}") ``` ### Using llm() for classification ```python - # Get document content - content = get_document("Q1 Report") - # Use llm() to classify sentiment - sentiment = llm(f"Classify the sentiment as positive, negative, or mixed: {content}") + content = await get_document("Q1 Report") + sentiment = await llm(f"Classify the sentiment as positive, negative, or mixed: {content}") print(sentiment) ``` @@ -548,23 +491,18 @@ interactions: role: system - content: How many documents are in the database? role: user - - content: |- - - We need to list documents. - + - content: null + reasoning: Need to list docs. role: assistant tool_calls: - function: - arguments: '{"code":"# list documents\nimport json\nprint(list_documents())\n"}' + arguments: '{"code":"docs = await list_documents(limit=1000)\nprint(len(docs))"}' name: execute_code - id: call_d8xhmimu + id: call_stp0fimx type: function - - content: '{"code":"# list documents\nimport json\nprint(list_documents())\n","stdout":"[{''id'': ''b73f8a17-4328-475c-84db-3d81ce52adce'', - ''title'': ''Doc 1'', ''uri'': None, ''created_at'': ''2026-02-06 11:21:55.843558''}, {''id'': ''accb877b-f04e-4bf2-ba4c-2d90339fa875'', - ''title'': ''Doc 2'', ''uri'': None, ''created_at'': ''2026-02-06 11:21:57.397026''}, {''id'': ''afdb966f-5e9d-4759-a08f-28eb5108c80f'', - ''title'': ''Doc 3'', ''uri'': None, ''created_at'': ''2026-02-06 11:21:58.988378''}]\n","stderr":"","success":true}' + - content: '{"code":"docs = await list_documents(limit=1000)\nprint(len(docs))","stdout":"3\n","stderr":"","success":true}' role: tool - tool_call_id: call_d8xhmimu + tool_call_id: call_stp0fimx model: gpt-oss reasoning_effort: low stream: false @@ -572,11 +510,10 @@ interactions: tools: - function: description: |- - Execute Python code in a Docker-sandboxed environment. + Execute Python code in a sandboxed interpreter. The code has access to haiku.rag functions (search, list_documents, - get_document, get_docling_document, llm) and any Python standard - library module. + get_document, get_chunk, llm). Use print() to output results. @@ -616,7 +553,7 @@ interactions: response: headers: content-length: - - '523' + - '416' content-type: - application/json parsed_body: @@ -624,19 +561,17 @@ interactions: - finish_reason: stop index: 0 message: - content: '{"answer":"There are 3 documents in the database.","program":"# List and count documents\nimport json\n\ndocs - = list_documents()\nprint(f\"Number of documents: {len(docs)}\")\n"}' - reasoning: Count is 3. Provide answer. + content: '{"answer":"There are 3 documents in the database.","program":"docs = await list_documents(limit=1000)\nprint(len(docs))"}' role: assistant - created: 1770373336 - id: chatcmpl-441 + created: 1771924498 + id: chatcmpl-945 model: gpt-oss object: chat.completion system_fingerprint: fp_ollama usage: - completion_tokens: 68 - prompt_tokens: 2019 - total_tokens: 2087 + completion_tokens: 38 + prompt_tokens: 1709 + total_tokens: 1747 status: code: 200 message: OK diff --git a/tests/cassettes/test_rlm/TestClientRLMIntegration.test_rlm_docling_document_structure.yaml b/tests/cassettes/test_rlm/TestClientRLMIntegration.test_rlm_docling_document_structure.yaml deleted file mode 100644 index 69e8a13e..00000000 --- a/tests/cassettes/test_rlm/TestClientRLMIntegration.test_rlm_docling_document_structure.yaml +++ /dev/null @@ -1,988 +0,0 @@ -interactions: -- request: - headers: - accept: - - application/json - accept-encoding: - - gzip, deflate, zstd - connection: - - keep-alive - content-length: - - '10466' - content-type: - - application/json - host: - - localhost:11434 - method: POST - parsed_body: - encoding_format: base64 - input: - - |2- - - Table 1: DocLayNet dataset overview. Along with the frequency of each class label, we present the relative occurrence (as % of row "Total") in the train, test and validation sets. The inter-annotator agreement is computed as the mAP@0.5-0.95 metric between pairwise annotations from the triple-annotated pages, from which we obtain accuracy ranges. - - Caption, Count = 22524. Caption, % of Total.Train = 2.04. Caption, % of Total.Test = 1.77. Caption, % of Total.Val - = 2.32. Caption, triple inter-annotator mAP @ 0.5-0.95 (%).All = 84-89. Caption, triple inter-annotator mAP @ 0.5-0.95 - (%).Fin = 40-61. Caption, triple inter-annotator mAP @ 0.5-0.95 (%).Man = 86-92. Caption, triple inter-annotator mAP - @ 0.5-0.95 (%).Sci = 94-99. Caption, triple inter-annotator mAP @ 0.5-0.95 (%).Law = 95-99. Caption, triple inter-annotator - mAP @ 0.5-0.95 (%).Pat = 69-78. Caption, triple inter-annotator mAP @ 0.5-0.95 (%).Ten = - - n/a. Footnote, Count = 6318. Footnote, % of Total.Train = 0.60. Footnote, % of Total.Test = 0.31. Footnote, % of Total.Val - = 0.58. Footnote, triple inter-annotator mAP @ 0.5-0.95 (%).All = 83-91. Footnote, triple inter-annotator mAP @ 0.5-0.95 - (%).Fin = n/a. Footnote, triple inter-annotator mAP @ 0.5-0.95 (%).Man = 100. Footnote, triple inter-annotator mAP - @ 0.5-0.95 (%).Sci = 62-88. Footnote, triple inter-annotator mAP @ 0.5-0.95 (%).Law = 85-94. Footnote, triple inter-annotator - mAP @ 0.5-0.95 (%).Pat = n/a. Footnote, triple inter-annotator mAP @ 0.5-0.95 (%).Ten - - = 82-97. Formula, Count = 25027. Formula, % of Total.Train = 2.25. Formula, % of Total.Test = 1.90. Formula, % of - Total.Val = 2.96. Formula, triple inter-annotator mAP @ 0.5-0.95 (%).All = 83-85. Formula, triple inter-annotator - mAP @ 0.5-0.95 (%).Fin = . Formula, triple inter-annotator mAP @ 0.5-0.95 (%).Man = n/a. Formula, triple inter-annotator - mAP @ 0.5-0.95 (%).Sci = 84-87. Formula, triple inter-annotator mAP @ 0.5-0.95 (%).Law = 86-96. Formula, triple inter-annotator - mAP @ 0.5-0.95 (%).Pat = . Formula, triple inter-annotator mAP @ 0.5-0.95 (%).Ten = n/a. List-item, Count = - - 185660. List-item, % of Total.Train = 17.19. List-item, % of Total.Test = 13.34. List-item, % of Total.Val = 15.82. - List-item, triple inter-annotator mAP @ 0.5-0.95 (%).All = 87-88. List-item, triple inter-annotator mAP @ 0.5-0.95 - (%).Fin = 74-83. List-item, triple inter-annotator mAP @ 0.5-0.95 (%).Man = 90-92. List-item, triple inter-annotator - mAP @ 0.5-0.95 (%).Sci = 97-97. List-item, triple inter-annotator mAP @ 0.5-0.95 (%).Law = 81-85. List-item, triple - inter-annotator mAP @ 0.5-0.95 (%).Pat = 75-88. List-item, triple inter-annotator mAP @ - - 0.5-0.95 (%).Ten = 93-95. Page-footer, Count = 70878. Page-footer, % of Total.Train = 6.51. Page-footer, % of Total.Test - = 5.58. Page-footer, % of Total.Val = 6.00. Page-footer, triple inter-annotator mAP @ 0.5-0.95 (%).All = 93-94. Page-footer, - triple inter-annotator mAP @ 0.5-0.95 (%).Fin = 88-90. Page-footer, triple inter-annotator mAP @ 0.5-0.95 (%).Man - = 95-96. Page-footer, triple inter-annotator mAP @ 0.5-0.95 (%).Sci = 100. Page-footer, triple inter-annotator mAP - @ 0.5-0.95 (%).Law = 92-97. Page-footer, triple inter-annotator mAP @ 0.5-0.95 (%).Pat = 100. - - Page-footer, triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 96-98. Page-header, Count = 58022. Page-header, % of - Total.Train = 5.10. Page-header, % of Total.Test = 6.70. Page-header, % of Total.Val = 5.06. Page-header, triple inter-annotator - mAP @ 0.5-0.95 (%).All = 85-89. Page-header, triple inter-annotator mAP @ 0.5-0.95 (%).Fin = 66-76. Page-header, triple - inter-annotator mAP @ 0.5-0.95 (%).Man = 90-94. Page-header, triple inter-annotator mAP @ 0.5-0.95 (%).Sci = 98-100. - Page-header, triple inter-annotator mAP @ 0.5-0.95 (%).Law = 91-92. Page-header, triple inter-annotator mAP @ - - 0.5-0.95 (%).Pat = 97-99. Page-header, triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 81-86. Picture, Count = 45976. - Picture, % of Total.Train = 4.21. Picture, % of Total.Test = 2.78. Picture, % of Total.Val = 5.31. Picture, triple - inter-annotator mAP @ 0.5-0.95 (%).All = 69-71. Picture, triple inter-annotator mAP @ 0.5-0.95 (%).Fin = 56-59. Picture, - triple inter-annotator mAP @ 0.5-0.95 (%).Man = 82-86. Picture, triple inter-annotator mAP @ 0.5-0.95 (%).Sci = 69-82. - Picture, triple inter-annotator mAP @ 0.5-0.95 (%).Law = 80-95. Picture, triple - - inter-annotator mAP @ 0.5-0.95 (%).Pat = 66-71. Picture, triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 59-76. Section-header, - Count = 142884. Section-header, % of Total.Train = 12.60. Section-header, % of Total.Test = 15.77. Section-header, - % of Total.Val = 12.85. Section-header, triple inter-annotator mAP @ 0.5-0.95 (%).All = 83-84. Section-header, triple - inter-annotator mAP @ 0.5-0.95 (%).Fin = 76-81. Section-header, triple inter-annotator mAP @ 0.5-0.95 (%).Man = 90-92. - Section-header, triple inter-annotator mAP @ 0.5-0.95 (%).Sci = 94-95. Section-header, triple inter-annotator mAP - @ - - 0.5-0.95 (%).Law = 87-94. Section-header, triple inter-annotator mAP @ 0.5-0.95 (%).Pat = 69-73. Section-header, triple - inter-annotator mAP @ 0.5-0.95 (%).Ten = 78-86. Table, Count = 34733. Table, % of Total.Train = 3.20. Table, % of - Total.Test = 2.27. Table, % of Total.Val = 3.60. Table, triple inter-annotator mAP @ 0.5-0.95 (%).All = 77-81. Table, - triple inter-annotator mAP @ 0.5-0.95 (%).Fin = 75-80. Table, triple inter-annotator mAP @ 0.5-0.95 (%).Man = 83-86. - Table, triple inter-annotator mAP @ 0.5-0.95 (%).Sci = 98-99. Table, triple - - inter-annotator mAP @ 0.5-0.95 (%).Law = 58-80. Table, triple inter-annotator mAP @ 0.5-0.95 (%).Pat = 79-84. Table, - triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 70-85. Text, Count = 510377. Text, % of Total.Train = 45.82. Text, - % of Total.Test = 49.28. Text, % of Total.Val = 45.00. Text, triple inter-annotator mAP @ 0.5-0.95 (%).All = 84-86. - Text, triple inter-annotator mAP @ 0.5-0.95 (%).Fin = 81-86. Text, triple inter-annotator mAP @ 0.5-0.95 (%).Man = - 88-93. Text, triple inter-annotator mAP @ 0.5-0.95 (%).Sci = - - 89-93. Text, triple inter-annotator mAP @ 0.5-0.95 (%).Law = 87-92. Text, triple inter-annotator mAP @ 0.5-0.95 (%).Pat - = 71-79. Text, triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 87-95. Title, Count = 5071. Title, % of Total.Train - = 0.47. Title, % of Total.Test = 0.30. Title, % of Total.Val = 0.50. Title, triple inter-annotator mAP @ 0.5-0.95 - (%).All = 60-72. Title, triple inter-annotator mAP @ 0.5-0.95 (%).Fin = 24-63. Title, triple inter-annotator mAP @ - 0.5-0.95 (%).Man = 50-63. Title, triple inter-annotator mAP @ 0.5-0.95 - - (%).Sci = 94-100. Title, triple inter-annotator mAP @ 0.5-0.95 (%).Law = 82-96. Title, triple inter-annotator mAP - @ 0.5-0.95 (%).Pat = 68-79. Title, triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 24-56. Total, Count = 1107470. - Total, % of Total.Train = 941123. Total, % of Total.Test = 99816. Total, % of Total.Val = 66531. Total, triple inter-annotator - mAP @ 0.5-0.95 (%).All = 82-83. Total, triple inter-annotator mAP @ 0.5-0.95 (%).Fin = 71-74. Total, triple inter-annotator - mAP @ 0.5-0.95 (%).Man = 79-81. Total, triple inter-annotator - - |- - mAP @ 0.5-0.95 (%).Sci = 89-94. Total, triple inter-annotator mAP @ 0.5-0.95 (%).Law = 86-91. Total, triple inter-annotator mAP @ 0.5-0.95 (%).Pat = 71-76. Total, triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 68-85 - Figure 3: Corpus Conversion Service annotation user interface. The PDF page is shown in the background, with overlaid text-cells (in darker shades). The annotation boxes can be drawn by dragging a rectangle over each segment with the respective label from the palette on the right. - we distributed the annotation workload and performed continuous quality controls. Phase one and two required a small team of experts only. For phases three and four, a group of 40 dedicated annotators were assembled and supervised. - - 'Phase 1: Data selection and preparation. Our inclusion criteria for documents were described in Section 3. A large - effort went into ensuring that all documents are free to use. The data sources include publication repositories such - as arXiv$^{3}$, government offices, company websites as well as data directory services for financial reports and - patents. Scanned documents were excluded wherever possible because they can be rotated or skewed. This would not allow - us to perform annotation with rectangular bounding-boxes and therefore complicate the annotation process.' - - 'Phase 2: Label selection and guideline. We reviewed the collected documents and identified the most common structural - features they exhibit. This was achieved by identifying recurrent layout elements and lead us to the definition of - 11 distinct class labels. These 11 class labels are $_{Caption}$, $_{Footnote}$, $_{Formula}$, $_{List-item}$, Page-$_{footer}$, - $_{Page-header}$, $_{Picture}$, $_{Section-header}$, $_{Table}$, $_{Text}$, and $_{Title}$. Critical factors that - were considered for the choice of these class labels were (1) the overall occurrence of the label, (2) the specificity - of the label, (3) recognisability on a single page (i.e. no need for context from previous or next page) and (4) overall - coverage of the page. Specificity ensures that the choice of label is not ambiguous, while coverage ensures that all - meaningful items on a page can be annotated. We refrained from class labels that are very specific to a document category, - such as Abstract in the Scientific Articles category. We also avoided class labels that are tightly linked to the - semantics of the text. Labels such as Author and' - - |- - $_{Affiliation}$, as seen in DocBank, are often only distinguishable by discriminating on - Preparation work included uploading and parsing the sourced PDF documents in the Corpus Conversion Service (CCS) [22], a cloud-native platform which provides a visual annotation interface and allows for dataset inspection and analysis. The annotation interface of CCS is shown in Figure 3. The desired balance of pages between the different document categories was achieved by selective subsampling of pages with certain desired properties. For example, we made sure to include the title page of each document and bias the remaining page selection to those with figures or tables. The latter was achieved by leveraging pre-trained object detection models from PubLayNet, which helped us estimate how many figures and tables a given page contains. - $^{3}$https://arxiv.org/ - model: qwen3-embedding:4b - uri: http://localhost:11434/v1/embeddings - response: - headers: - content-type: - - application/json - transfer-encoding: - - chunked - parsed_body: - data: - - embedding: 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 - index: 0 - object: embedding - - embedding: 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 - index: 1 - object: embedding - - embedding: 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 - index: 2 - object: embedding - - embedding: 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 - index: 3 - object: embedding - - embedding: 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 - index: 4 - object: embedding - - embedding: 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 - index: 5 - object: embedding - - embedding: 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 - index: 6 - object: embedding - - embedding: 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 - index: 7 - object: embedding - - embedding: 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 - index: 8 - object: embedding - - embedding: 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 - index: 9 - object: embedding - - embedding: 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 - index: 10 - object: embedding - - embedding: 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 - index: 11 - object: embedding - - embedding: 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 - index: 12 - object: embedding - - embedding: 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 - index: 13 - object: embedding - - embedding: 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 - index: 14 - object: embedding - - embedding: 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 - index: 15 - object: embedding - - embedding: 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 - index: 16 - object: embedding - model: qwen3-embedding:4b - object: list - usage: - prompt_tokens: 3883 - total_tokens: 3883 - status: - code: 200 - message: OK -- request: - headers: - accept: - - application/json - accept-encoding: - - gzip, deflate, zstd - connection: - - keep-alive - content-length: - - '7823' - content-type: - - application/json - host: - - localhost:11434 - method: POST - parsed_body: - messages: - - content: |- - You are a Recursive Language Model (RLM) agent that solves complex research questions by writing and executing Python code. - - IMPORTANT: You MUST use the `execute_code` tool to run Python code. The functions described below are ONLY available inside the execute_code tool - you cannot access them any other way. Always execute code to answer questions; do not just describe what code would do. - - CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just use them directly: - - search("query") ✓ CORRECT - - from haiku.rag import search ✗ WRONG - will fail - - You have access to a sandboxed Python environment with these haiku.rag functions (use them directly, no imports needed): - - ## Available Functions - - ### search(query, limit=10) -> list[dict] - Search the knowledge base using hybrid search (vector + full-text). - Returns list of dicts with keys: chunk_id, content, document_id, document_title, document_uri, score, page_numbers, headings - - ### list_documents(limit=10, offset=0) -> list[dict] - List available documents in the knowledge base. - Returns list of dicts with keys: id, title, uri, created_at - - ### get_document(id_or_title) -> str | None - Get the full text content of a document by ID, title, or URI. - Returns the document content as a string, or None if not found. - - ### get_docling_document(id_or_title) -> DoclingDocument | None - Get the structured DoclingDocument object for advanced analysis. - Returns a DoclingDocument object, or None if not found. - See "DoclingDocument API" section below for how to use it. - - ### llm(prompt) -> str - Call an LLM directly with the given prompt. Returns the response as a string. - Use this for classification, summarization, extraction, or any task where you - already have the content and just need LLM reasoning. - - ## Pre-loaded Documents Variable - - If documents were pre-loaded for this session, a `documents` variable is available: - ```python - # documents is a list of dicts with keys: id, title, uri, content - for doc in documents: - print(doc['title'], len(doc['content'])) - ``` - Check if it exists with: `if 'documents' in dir(): ...` - - ## Standard Library Modules - You can import any Python standard library module. - - ## Strategy Guide - - 1. **Explore First**: Start by listing documents or searching to understand what's available. Document names may differ from filenames (e.g., "tbmed593.pdf" might be stored as "TB MED 593" or similar). - 2. **If get_document returns None**: Use `list_documents()` to see actual document titles, or `search()` to find relevant content. - 3. **Iterative Refinement**: Run code, examine results, adjust your approach based on what you find. - 4. **Use print() Liberally**: The REPL captures stdout - print intermediate results to see what you're working with. - 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and use collections. - 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use llm(). - 7. **Cite Your Sources**: Track which documents/chunks informed your answer for citation. - - ## DoclingDocument API - - When you call `get_docling_document(id_or_title)`, you get a DoclingDocument object for structured document analysis. - - ### Properties - - `doc.texts` - List of all text items (paragraphs, headings, etc.) - - `doc.tables` - List of all tables - - `doc.pictures` - List of all pictures/figures - - `doc.name` - Document name - - ### Methods - - `doc.iterate_items(with_groups=False)` - Iterate all items with hierarchy level - Returns tuples of (item, level) where level is nesting depth - - `doc.export_to_markdown()` - Export entire document as markdown string - - ### Text Item Properties - - `item.text` - The text content - - `item.label` - Type: title, paragraph, section_header, list_item, etc. (lowercase enum values) - - `item.prov` - Provenance (page numbers, bounding boxes) - - ### Table Access - - `table.data.num_rows`, `table.data.num_cols` - Dimensions - - `table.data.table_cells` - List of TableCell objects - - `cell.text`, `cell.start_row_offset_idx`, `cell.start_col_offset_idx` - - ### Example Usage - ```python - doc = get_docling_document("My Document") - - # Get all headings - headings = [t.text for t in doc.texts if "header" in str(t.label)] - - # Iterate with structure - for item, level in doc.iterate_items(): - print(" " * level + item.text[:50]) - - # Extract table data - for table in doc.tables: - for cell in table.data.table_cells: - print(f"Row {cell.start_row_offset_idx}, Col {cell.start_col_offset_idx}: {cell.text}") - ``` - - ## Example Patterns - - ### Counting documents matching a condition - ```python - docs = list_documents(limit=100) - count = 0 - for doc in docs: - content = get_document(doc['id']) - if content and 'keyword' in content.lower(): - count += 1 - print(f"Found in: {doc['title']}") - print(f"Total: {count}") - ``` - - ### Aggregating data across documents - ```python - import re - numbers = [] - results = search("financial data", limit=20) - for r in results: - matches = re.findall(r'\$([\d,]+)', r['content']) - for m in matches: - numbers.append(int(m.replace(',', ''))) - print(f"Average: ${sum(numbers)/len(numbers):,.2f}") - ``` - - ### Using llm() for classification - ```python - # Get document content - content = get_document("Q1 Report") - # Use llm() to classify sentiment - sentiment = llm(f"Classify the sentiment as positive, negative, or mixed: {content}") - print(sentiment) - ``` - - ## Workflow - - 1. **ALWAYS start by using execute_code** to explore the knowledge base - 2. Run multiple code blocks as needed to gather information - 3. After collecting data, provide your final answer - - ## Output Format - - CRITICAL: Your final response MUST be valid JSON matching this exact schema: - ```json - {"answer": "Your complete answer here as a string", "program": "Your final consolidated program here as a string"} - ``` - - - `answer`: A clear answer to the user's question with key findings and references to specific documents/chunks. - - `program`: A single, self-contained Python program that produces the answer. Consolidate your exploratory code executions into one clean script. - - Do NOT return arbitrary JSON structures. Always use the exact format: {"answer": "...", "program": "..."} - - CRITICAL: You MUST call execute_code at least once before providing your answer. Never give up without trying to execute code first. - role: system - - content: How many tables are in the document? Also tell me how many pictures/figures it contains. - role: user - model: gpt-oss - reasoning_effort: low - stream: false - tool_choice: auto - tools: - - function: - description: |- - Execute Python code in a Docker-sandboxed environment. - - The code has access to haiku.rag functions (search, list_documents, - get_document, get_docling_document, llm) and any Python standard - library module. - - Use print() to output results. - - Structured result with success status, stdout, and stderr. - - name: execute_code - parameters: - additionalProperties: false - properties: - code: - description: Python code to execute. - type: string - required: - - code - type: object - strict: true - type: function - - function: - description: Result from RLM agent execution. - name: final_result - parameters: - additionalProperties: false - properties: - answer: - description: The answer to the user's question - type: string - program: - description: The final consolidated program - type: string - required: - - answer - - program - type: object - strict: true - type: function - uri: http://localhost:11434/v1/chat/completions - response: - headers: - content-length: - - '589' - content-type: - - application/json - parsed_body: - choices: - - finish_reason: tool_calls - index: 0 - message: - content: '' - reasoning: We need to find a document. Likely the user refers to a specific document loaded. Let's list documents. - role: assistant - tool_calls: - - function: - arguments: '{"code":"docs = list_documents(limit=10)\nprint(docs)\n"}' - name: execute_code - id: call_qmo5y6dt - index: 0 - type: function - created: 1770373391 - id: chatcmpl-318 - model: gpt-oss - object: chat.completion - system_fingerprint: fp_ollama - usage: - completion_tokens: 59 - prompt_tokens: 1759 - total_tokens: 1818 - status: - code: 200 - message: OK -- request: - headers: - accept: - - application/json - accept-encoding: - - gzip, deflate, zstd - connection: - - keep-alive - content-length: - - '8531' - content-type: - - application/json - host: - - localhost:11434 - method: POST - parsed_body: - messages: - - content: |- - You are a Recursive Language Model (RLM) agent that solves complex research questions by writing and executing Python code. - - IMPORTANT: You MUST use the `execute_code` tool to run Python code. The functions described below are ONLY available inside the execute_code tool - you cannot access them any other way. Always execute code to answer questions; do not just describe what code would do. - - CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just use them directly: - - search("query") ✓ CORRECT - - from haiku.rag import search ✗ WRONG - will fail - - You have access to a sandboxed Python environment with these haiku.rag functions (use them directly, no imports needed): - - ## Available Functions - - ### search(query, limit=10) -> list[dict] - Search the knowledge base using hybrid search (vector + full-text). - Returns list of dicts with keys: chunk_id, content, document_id, document_title, document_uri, score, page_numbers, headings - - ### list_documents(limit=10, offset=0) -> list[dict] - List available documents in the knowledge base. - Returns list of dicts with keys: id, title, uri, created_at - - ### get_document(id_or_title) -> str | None - Get the full text content of a document by ID, title, or URI. - Returns the document content as a string, or None if not found. - - ### get_docling_document(id_or_title) -> DoclingDocument | None - Get the structured DoclingDocument object for advanced analysis. - Returns a DoclingDocument object, or None if not found. - See "DoclingDocument API" section below for how to use it. - - ### llm(prompt) -> str - Call an LLM directly with the given prompt. Returns the response as a string. - Use this for classification, summarization, extraction, or any task where you - already have the content and just need LLM reasoning. - - ## Pre-loaded Documents Variable - - If documents were pre-loaded for this session, a `documents` variable is available: - ```python - # documents is a list of dicts with keys: id, title, uri, content - for doc in documents: - print(doc['title'], len(doc['content'])) - ``` - Check if it exists with: `if 'documents' in dir(): ...` - - ## Standard Library Modules - You can import any Python standard library module. - - ## Strategy Guide - - 1. **Explore First**: Start by listing documents or searching to understand what's available. Document names may differ from filenames (e.g., "tbmed593.pdf" might be stored as "TB MED 593" or similar). - 2. **If get_document returns None**: Use `list_documents()` to see actual document titles, or `search()` to find relevant content. - 3. **Iterative Refinement**: Run code, examine results, adjust your approach based on what you find. - 4. **Use print() Liberally**: The REPL captures stdout - print intermediate results to see what you're working with. - 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and use collections. - 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use llm(). - 7. **Cite Your Sources**: Track which documents/chunks informed your answer for citation. - - ## DoclingDocument API - - When you call `get_docling_document(id_or_title)`, you get a DoclingDocument object for structured document analysis. - - ### Properties - - `doc.texts` - List of all text items (paragraphs, headings, etc.) - - `doc.tables` - List of all tables - - `doc.pictures` - List of all pictures/figures - - `doc.name` - Document name - - ### Methods - - `doc.iterate_items(with_groups=False)` - Iterate all items with hierarchy level - Returns tuples of (item, level) where level is nesting depth - - `doc.export_to_markdown()` - Export entire document as markdown string - - ### Text Item Properties - - `item.text` - The text content - - `item.label` - Type: title, paragraph, section_header, list_item, etc. (lowercase enum values) - - `item.prov` - Provenance (page numbers, bounding boxes) - - ### Table Access - - `table.data.num_rows`, `table.data.num_cols` - Dimensions - - `table.data.table_cells` - List of TableCell objects - - `cell.text`, `cell.start_row_offset_idx`, `cell.start_col_offset_idx` - - ### Example Usage - ```python - doc = get_docling_document("My Document") - - # Get all headings - headings = [t.text for t in doc.texts if "header" in str(t.label)] - - # Iterate with structure - for item, level in doc.iterate_items(): - print(" " * level + item.text[:50]) - - # Extract table data - for table in doc.tables: - for cell in table.data.table_cells: - print(f"Row {cell.start_row_offset_idx}, Col {cell.start_col_offset_idx}: {cell.text}") - ``` - - ## Example Patterns - - ### Counting documents matching a condition - ```python - docs = list_documents(limit=100) - count = 0 - for doc in docs: - content = get_document(doc['id']) - if content and 'keyword' in content.lower(): - count += 1 - print(f"Found in: {doc['title']}") - print(f"Total: {count}") - ``` - - ### Aggregating data across documents - ```python - import re - numbers = [] - results = search("financial data", limit=20) - for r in results: - matches = re.findall(r'\$([\d,]+)', r['content']) - for m in matches: - numbers.append(int(m.replace(',', ''))) - print(f"Average: ${sum(numbers)/len(numbers):,.2f}") - ``` - - ### Using llm() for classification - ```python - # Get document content - content = get_document("Q1 Report") - # Use llm() to classify sentiment - sentiment = llm(f"Classify the sentiment as positive, negative, or mixed: {content}") - print(sentiment) - ``` - - ## Workflow - - 1. **ALWAYS start by using execute_code** to explore the knowledge base - 2. Run multiple code blocks as needed to gather information - 3. After collecting data, provide your final answer - - ## Output Format - - CRITICAL: Your final response MUST be valid JSON matching this exact schema: - ```json - {"answer": "Your complete answer here as a string", "program": "Your final consolidated program here as a string"} - ``` - - - `answer`: A clear answer to the user's question with key findings and references to specific documents/chunks. - - `program`: A single, self-contained Python program that produces the answer. Consolidate your exploratory code executions into one clean script. - - Do NOT return arbitrary JSON structures. Always use the exact format: {"answer": "...", "program": "..."} - - CRITICAL: You MUST call execute_code at least once before providing your answer. Never give up without trying to execute code first. - role: system - - content: How many tables are in the document? Also tell me how many pictures/figures it contains. - role: user - - content: |- - - We need to find a document. Likely the user refers to a specific document loaded. Let's list documents. - - role: assistant - tool_calls: - - function: - arguments: '{"code":"docs = list_documents(limit=10)\nprint(docs)\n"}' - name: execute_code - id: call_qmo5y6dt - type: function - - content: '{"code":"docs = list_documents(limit=10)\nprint(docs)\n","stdout":"[{''id'': ''8b7e395a-b2b4-423d-8c65-53d7ea800da1'', - ''title'': None, ''uri'': ''file:///Users/ggozad/dev/open-source/haiku.rag-feat-recursive-llm/tests/data/doclaynet.pdf'', - ''created_at'': ''2026-02-06 11:23:09.536783''}]\n","stderr":"","success":true}' - role: tool - tool_call_id: call_qmo5y6dt - model: gpt-oss - reasoning_effort: low - stream: false - tool_choice: auto - tools: - - function: - description: |- - Execute Python code in a Docker-sandboxed environment. - - The code has access to haiku.rag functions (search, list_documents, - get_document, get_docling_document, llm) and any Python standard - library module. - - Use print() to output results. - - Structured result with success status, stdout, and stderr. - - name: execute_code - parameters: - additionalProperties: false - properties: - code: - description: Python code to execute. - type: string - required: - - code - type: object - strict: true - type: function - - function: - description: Result from RLM agent execution. - name: final_result - parameters: - additionalProperties: false - properties: - answer: - description: The answer to the user's question - type: string - program: - description: The final consolidated program - type: string - required: - - answer - - program - type: object - strict: true - type: function - uri: http://localhost:11434/v1/chat/completions - response: - headers: - content-length: - - '691' - content-type: - - application/json - parsed_body: - choices: - - finish_reason: tool_calls - index: 0 - message: - content: '' - reasoning: List shows one doc. Let's get docling. - role: assistant - tool_calls: - - function: - arguments: '{"code":"doc = get_docling_document(''file:///Users/ggozad/dev/open-source/haiku.rag-feat-recursive-llm/tests/data/doclaynet.pdf'')\nprint(doc.name)\nprint(''tables'', - len(doc.tables))\nprint(''pictures'', len(doc.pictures))\n"}' - name: execute_code - id: call_q4w5nsrp - index: 0 - type: function - created: 1770373393 - id: chatcmpl-498 - model: gpt-oss - object: chat.completion - system_fingerprint: fp_ollama - usage: - completion_tokens: 96 - prompt_tokens: 1953 - total_tokens: 2049 - status: - code: 200 - message: OK -- request: - headers: - accept: - - application/json - accept-encoding: - - gzip, deflate, zstd - connection: - - keep-alive - content-length: - - '9333' - content-type: - - application/json - host: - - localhost:11434 - method: POST - parsed_body: - messages: - - content: |- - You are a Recursive Language Model (RLM) agent that solves complex research questions by writing and executing Python code. - - IMPORTANT: You MUST use the `execute_code` tool to run Python code. The functions described below are ONLY available inside the execute_code tool - you cannot access them any other way. Always execute code to answer questions; do not just describe what code would do. - - CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just use them directly: - - search("query") ✓ CORRECT - - from haiku.rag import search ✗ WRONG - will fail - - You have access to a sandboxed Python environment with these haiku.rag functions (use them directly, no imports needed): - - ## Available Functions - - ### search(query, limit=10) -> list[dict] - Search the knowledge base using hybrid search (vector + full-text). - Returns list of dicts with keys: chunk_id, content, document_id, document_title, document_uri, score, page_numbers, headings - - ### list_documents(limit=10, offset=0) -> list[dict] - List available documents in the knowledge base. - Returns list of dicts with keys: id, title, uri, created_at - - ### get_document(id_or_title) -> str | None - Get the full text content of a document by ID, title, or URI. - Returns the document content as a string, or None if not found. - - ### get_docling_document(id_or_title) -> DoclingDocument | None - Get the structured DoclingDocument object for advanced analysis. - Returns a DoclingDocument object, or None if not found. - See "DoclingDocument API" section below for how to use it. - - ### llm(prompt) -> str - Call an LLM directly with the given prompt. Returns the response as a string. - Use this for classification, summarization, extraction, or any task where you - already have the content and just need LLM reasoning. - - ## Pre-loaded Documents Variable - - If documents were pre-loaded for this session, a `documents` variable is available: - ```python - # documents is a list of dicts with keys: id, title, uri, content - for doc in documents: - print(doc['title'], len(doc['content'])) - ``` - Check if it exists with: `if 'documents' in dir(): ...` - - ## Standard Library Modules - You can import any Python standard library module. - - ## Strategy Guide - - 1. **Explore First**: Start by listing documents or searching to understand what's available. Document names may differ from filenames (e.g., "tbmed593.pdf" might be stored as "TB MED 593" or similar). - 2. **If get_document returns None**: Use `list_documents()` to see actual document titles, or `search()` to find relevant content. - 3. **Iterative Refinement**: Run code, examine results, adjust your approach based on what you find. - 4. **Use print() Liberally**: The REPL captures stdout - print intermediate results to see what you're working with. - 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and use collections. - 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use llm(). - 7. **Cite Your Sources**: Track which documents/chunks informed your answer for citation. - - ## DoclingDocument API - - When you call `get_docling_document(id_or_title)`, you get a DoclingDocument object for structured document analysis. - - ### Properties - - `doc.texts` - List of all text items (paragraphs, headings, etc.) - - `doc.tables` - List of all tables - - `doc.pictures` - List of all pictures/figures - - `doc.name` - Document name - - ### Methods - - `doc.iterate_items(with_groups=False)` - Iterate all items with hierarchy level - Returns tuples of (item, level) where level is nesting depth - - `doc.export_to_markdown()` - Export entire document as markdown string - - ### Text Item Properties - - `item.text` - The text content - - `item.label` - Type: title, paragraph, section_header, list_item, etc. (lowercase enum values) - - `item.prov` - Provenance (page numbers, bounding boxes) - - ### Table Access - - `table.data.num_rows`, `table.data.num_cols` - Dimensions - - `table.data.table_cells` - List of TableCell objects - - `cell.text`, `cell.start_row_offset_idx`, `cell.start_col_offset_idx` - - ### Example Usage - ```python - doc = get_docling_document("My Document") - - # Get all headings - headings = [t.text for t in doc.texts if "header" in str(t.label)] - - # Iterate with structure - for item, level in doc.iterate_items(): - print(" " * level + item.text[:50]) - - # Extract table data - for table in doc.tables: - for cell in table.data.table_cells: - print(f"Row {cell.start_row_offset_idx}, Col {cell.start_col_offset_idx}: {cell.text}") - ``` - - ## Example Patterns - - ### Counting documents matching a condition - ```python - docs = list_documents(limit=100) - count = 0 - for doc in docs: - content = get_document(doc['id']) - if content and 'keyword' in content.lower(): - count += 1 - print(f"Found in: {doc['title']}") - print(f"Total: {count}") - ``` - - ### Aggregating data across documents - ```python - import re - numbers = [] - results = search("financial data", limit=20) - for r in results: - matches = re.findall(r'\$([\d,]+)', r['content']) - for m in matches: - numbers.append(int(m.replace(',', ''))) - print(f"Average: ${sum(numbers)/len(numbers):,.2f}") - ``` - - ### Using llm() for classification - ```python - # Get document content - content = get_document("Q1 Report") - # Use llm() to classify sentiment - sentiment = llm(f"Classify the sentiment as positive, negative, or mixed: {content}") - print(sentiment) - ``` - - ## Workflow - - 1. **ALWAYS start by using execute_code** to explore the knowledge base - 2. Run multiple code blocks as needed to gather information - 3. After collecting data, provide your final answer - - ## Output Format - - CRITICAL: Your final response MUST be valid JSON matching this exact schema: - ```json - {"answer": "Your complete answer here as a string", "program": "Your final consolidated program here as a string"} - ``` - - - `answer`: A clear answer to the user's question with key findings and references to specific documents/chunks. - - `program`: A single, self-contained Python program that produces the answer. Consolidate your exploratory code executions into one clean script. - - Do NOT return arbitrary JSON structures. Always use the exact format: {"answer": "...", "program": "..."} - - CRITICAL: You MUST call execute_code at least once before providing your answer. Never give up without trying to execute code first. - role: system - - content: How many tables are in the document? Also tell me how many pictures/figures it contains. - role: user - - content: |- - - We need to find a document. Likely the user refers to a specific document loaded. Let's list documents. - - role: assistant - tool_calls: - - function: - arguments: '{"code":"docs = list_documents(limit=10)\nprint(docs)\n"}' - name: execute_code - id: call_qmo5y6dt - type: function - - content: '{"code":"docs = list_documents(limit=10)\nprint(docs)\n","stdout":"[{''id'': ''8b7e395a-b2b4-423d-8c65-53d7ea800da1'', - ''title'': None, ''uri'': ''file:///Users/ggozad/dev/open-source/haiku.rag-feat-recursive-llm/tests/data/doclaynet.pdf'', - ''created_at'': ''2026-02-06 11:23:09.536783''}]\n","stderr":"","success":true}' - role: tool - tool_call_id: call_qmo5y6dt - - content: |- - - List shows one doc. Let's get docling. - - role: assistant - tool_calls: - - function: - arguments: '{"code":"doc = get_docling_document(''file:///Users/ggozad/dev/open-source/haiku.rag-feat-recursive-llm/tests/data/doclaynet.pdf'')\nprint(doc.name)\nprint(''tables'', - len(doc.tables))\nprint(''pictures'', len(doc.pictures))\n"}' - name: execute_code - id: call_q4w5nsrp - type: function - - content: '{"code":"doc = get_docling_document(''file:///Users/ggozad/dev/open-source/haiku.rag-feat-recursive-llm/tests/data/doclaynet.pdf'')\nprint(doc.name)\nprint(''tables'', - len(doc.tables))\nprint(''pictures'', len(doc.pictures))\n","stdout":"doclaynet\ntables 1\npictures 1\n","stderr":"","success":true}' - role: tool - tool_call_id: call_q4w5nsrp - model: gpt-oss - reasoning_effort: low - stream: false - tool_choice: auto - tools: - - function: - description: |- - Execute Python code in a Docker-sandboxed environment. - - The code has access to haiku.rag functions (search, list_documents, - get_document, get_docling_document, llm) and any Python standard - library module. - - Use print() to output results. - - Structured result with success status, stdout, and stderr. - - name: execute_code - parameters: - additionalProperties: false - properties: - code: - description: Python code to execute. - type: string - required: - - code - type: object - strict: true - type: function - - function: - description: Result from RLM agent execution. - name: final_result - parameters: - additionalProperties: false - properties: - answer: - description: The answer to the user's question - type: string - program: - description: The final consolidated program - type: string - required: - - answer - - program - type: object - strict: true - type: function - uri: http://localhost:11434/v1/chat/completions - response: - headers: - content-length: - - '1173' - content-type: - - application/json - parsed_body: - choices: - - finish_reason: stop - index: 0 - message: - content: '{"answer":"The document contains **1 table** and **1 picture/figure**.\n\nSources:\n- Docling metadata - shows 1 table and 1 picture.", "program":"# Retrieve structured document and report table and picture counts\n\ndef - main():\n # The document URL (as known from list_documents)\n doc_uri = ''file:///Users/ggozad/dev/open-source/haiku.rag-feat-recursive-llm/tests/data/doclaynet.pdf''\n # - Load DoclingDocument\n doc = get_docling_document(doc_uri)\n if not doc:\n print(\"Document not found\")\n return\n # - Count tables and pictures\n table_count = len(doc.tables)\n picture_count = len(doc.pictures)\n # Output - results\n print(f\"Document: {doc.name}\")\n print(f\"Tables: {table_count}\")\n print(f\"Pictures/figures: - {picture_count}\")\n\nif __name__ == \"__main__\":\n main()\n"}' - role: assistant - created: 1770373399 - id: chatcmpl-510 - model: gpt-oss - object: chat.completion - system_fingerprint: fp_ollama - usage: - completion_tokens: 237 - prompt_tokens: 2155 - total_tokens: 2392 - status: - code: 200 - message: OK -version: 1 diff --git a/tests/cassettes/test_rlm/TestClientRLMIntegration.test_rlm_search_and_extract.yaml b/tests/cassettes/test_rlm/TestClientRLMIntegration.test_rlm_search_and_extract.yaml index e133dc9d..04962a11 100644 --- a/tests/cassettes/test_rlm/TestClientRLMIntegration.test_rlm_search_and_extract.yaml +++ b/tests/cassettes/test_rlm/TestClientRLMIntegration.test_rlm_search_and_extract.yaml @@ -182,7 +182,7 @@ interactions: connection: - keep-alive content-length: - - '7869' + - '7420' content-type: - application/json host: @@ -195,32 +195,33 @@ interactions: IMPORTANT: You MUST use the `execute_code` tool to run Python code. The functions described below are ONLY available inside the execute_code tool - you cannot access them any other way. Always execute code to answer questions; do not just describe what code would do. - CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just use them directly: - - search("query") ✓ CORRECT + CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just call them with `await`: + - results = await search("query") ✓ CORRECT - from haiku.rag import search ✗ WRONG - will fail + - results = search("query") ✗ WRONG - must use await - You have access to a sandboxed Python environment with these haiku.rag functions (use them directly, no imports needed): + You have access to a sandboxed Python interpreter with these haiku.rag functions (use them directly with `await`, no imports needed): ## Available Functions - ### search(query, limit=10) -> list[dict] + ### await search(query, limit=10) -> list[dict] Search the knowledge base using hybrid search (vector + full-text). Returns list of dicts with keys: chunk_id, content, document_id, document_title, document_uri, score, page_numbers, headings - ### list_documents(limit=10, offset=0) -> list[dict] + ### await list_documents(limit=10, offset=0) -> list[dict] List available documents in the knowledge base. Returns list of dicts with keys: id, title, uri, created_at - ### get_document(id_or_title) -> str | None + ### await get_document(id_or_title) -> str | None Get the full text content of a document by ID, title, or URI. Returns the document content as a string, or None if not found. - ### get_docling_document(id_or_title) -> DoclingDocument | None - Get the structured DoclingDocument object for advanced analysis. - Returns a DoclingDocument object, or None if not found. - See "DoclingDocument API" section below for how to use it. + ### await get_chunk(chunk_id) -> dict | None + Get a specific chunk by its ID (from search results). + Returns dict with keys: chunk_id, content, document_id, document_title, headings, page_numbers, labels + Use this to retrieve full chunk details and metadata for citation. - ### llm(prompt) -> str + ### await llm(prompt) -> str Call an LLM directly with the given prompt. Returns the response as a string. Use this for classification, summarization, extraction, or any task where you already have the content and just need LLM reasoning. @@ -235,93 +236,64 @@ interactions: ``` Check if it exists with: `if 'documents' in dir(): ...` - ## Standard Library Modules - You can import any Python standard library module. + ## Available Python Features + + The interpreter supports: variables, arithmetic, strings, f-strings, lists, dicts, tuples, sets, loops, conditionals, comprehensions, functions, async/await, `map()`, `sorted()`/`.sort(key=...)`, try/except, and the `json` module. + + Not supported: imports (other than `json`), class definitions, generators/yield, match statements, decorators, `with` statements. + + For pattern matching or text extraction, use string methods (`str.split`, `str.find`, `str.startswith`, `in` operator) or the `llm()` function. ## Strategy Guide 1. **Explore First**: Start by listing documents or searching to understand what's available. Document names may differ from filenames (e.g., "tbmed593.pdf" might be stored as "TB MED 593" or similar). - 2. **If get_document returns None**: Use `list_documents()` to see actual document titles, or `search()` to find relevant content. + 2. **If get_document returns None**: Use `await list_documents()` to see actual document titles, or `await search()` to find relevant content. 3. **Iterative Refinement**: Run code, examine results, adjust your approach based on what you find. - 4. **Use print() Liberally**: The REPL captures stdout - print intermediate results to see what you're working with. - 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and use collections. - 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use llm(). - 7. **Cite Your Sources**: Track which documents/chunks informed your answer for citation. - - ## DoclingDocument API - - When you call `get_docling_document(id_or_title)`, you get a DoclingDocument object for structured document analysis. - - ### Properties - - `doc.texts` - List of all text items (paragraphs, headings, etc.) - - `doc.tables` - List of all tables - - `doc.pictures` - List of all pictures/figures - - `doc.name` - Document name - - ### Methods - - `doc.iterate_items(with_groups=False)` - Iterate all items with hierarchy level - Returns tuples of (item, level) where level is nesting depth - - `doc.export_to_markdown()` - Export entire document as markdown string - - ### Text Item Properties - - `item.text` - The text content - - `item.label` - Type: title, paragraph, section_header, list_item, etc. (lowercase enum values) - - `item.prov` - Provenance (page numbers, bounding boxes) - - ### Table Access - - `table.data.num_rows`, `table.data.num_cols` - Dimensions - - `table.data.table_cells` - List of TableCell objects - - `cell.text`, `cell.start_row_offset_idx`, `cell.start_col_offset_idx` - - ### Example Usage - ```python - doc = get_docling_document("My Document") - - # Get all headings - headings = [t.text for t in doc.texts if "header" in str(t.label)] - - # Iterate with structure - for item, level in doc.iterate_items(): - print(" " * level + item.text[:50]) - - # Extract table data - for table in doc.tables: - for cell in table.data.table_cells: - print(f"Row {cell.start_row_offset_idx}, Col {cell.start_col_offset_idx}: {cell.text}") - ``` + 4. **Use print() Liberally**: The sandbox captures stdout - print intermediate results to see what you're working with. + 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and data structures. + 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use `await llm()`. + 7. **Cite Your Sources**: Use get_chunk() to retrieve chunk metadata for citations. Track which documents/chunks informed your answer. ## Example Patterns ### Counting documents matching a condition ```python - docs = list_documents(limit=100) + docs = await list_documents(limit=100) count = 0 for doc in docs: - content = get_document(doc['id']) + content = await get_document(doc['id']) if content and 'keyword' in content.lower(): count += 1 print(f"Found in: {doc['title']}") print(f"Total: {count}") ``` - ### Aggregating data across documents + ### Extracting data with llm() ```python - import re numbers = [] - results = search("financial data", limit=20) + results = await search("financial data", limit=20) for r in results: - matches = re.findall(r'\$([\d,]+)', r['content']) - for m in matches: - numbers.append(int(m.replace(',', ''))) - print(f"Average: ${sum(numbers)/len(numbers):,.2f}") + extracted = await llm(f"Extract all dollar amounts from this text as a comma-separated list of numbers (no $ signs): {r['content']}") + for part in extracted.split(','): + part = part.strip().replace(',', '') + if part.isdigit(): + numbers.append(int(part)) + if numbers: + print(f"Average: {sum(numbers) / len(numbers)}") + ``` + + ### Using search results with get_chunk for citations + ```python + results = await search("safety requirements", limit=5) + for r in results: + chunk = await get_chunk(r['chunk_id']) + print(f"From '{chunk['document_title']}', page {chunk['page_numbers']}: {chunk['content'][:100]}") ``` ### Using llm() for classification ```python - # Get document content - content = get_document("Q1 Report") - # Use llm() to classify sentiment - sentiment = llm(f"Classify the sentiment as positive, negative, or mixed: {content}") + content = await get_document("Q1 Report") + sentiment = await llm(f"Classify the sentiment as positive, negative, or mixed: {content}") print(sentiment) ``` @@ -355,11 +327,10 @@ interactions: tools: - function: description: |- - Execute Python code in a Docker-sandboxed environment. + Execute Python code in a sandboxed interpreter. The code has access to haiku.rag functions (search, list_documents, - get_document, get_docling_document, llm) and any Python standard - library module. + get_document, get_chunk, llm). Use print() to output results. @@ -399,246 +370,7 @@ interactions: response: headers: content-length: - - '199' - content-type: - - application/json - parsed_body: - error: - code: null - message: 'error parsing tool call: raw=''search("document element types labels")'', err=invalid character ''s'' looking - for beginning of value' - param: null - type: api_error - status: - code: 500 - message: Internal Server Error -- request: - headers: - accept: - - application/json - accept-encoding: - - gzip, deflate, zstd - connection: - - keep-alive - content-length: - - '7869' - content-type: - - application/json - host: - - localhost:11434 - method: POST - parsed_body: - messages: - - content: |- - You are a Recursive Language Model (RLM) agent that solves complex research questions by writing and executing Python code. - - IMPORTANT: You MUST use the `execute_code` tool to run Python code. The functions described below are ONLY available inside the execute_code tool - you cannot access them any other way. Always execute code to answer questions; do not just describe what code would do. - - CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just use them directly: - - search("query") ✓ CORRECT - - from haiku.rag import search ✗ WRONG - will fail - - You have access to a sandboxed Python environment with these haiku.rag functions (use them directly, no imports needed): - - ## Available Functions - - ### search(query, limit=10) -> list[dict] - Search the knowledge base using hybrid search (vector + full-text). - Returns list of dicts with keys: chunk_id, content, document_id, document_title, document_uri, score, page_numbers, headings - - ### list_documents(limit=10, offset=0) -> list[dict] - List available documents in the knowledge base. - Returns list of dicts with keys: id, title, uri, created_at - - ### get_document(id_or_title) -> str | None - Get the full text content of a document by ID, title, or URI. - Returns the document content as a string, or None if not found. - - ### get_docling_document(id_or_title) -> DoclingDocument | None - Get the structured DoclingDocument object for advanced analysis. - Returns a DoclingDocument object, or None if not found. - See "DoclingDocument API" section below for how to use it. - - ### llm(prompt) -> str - Call an LLM directly with the given prompt. Returns the response as a string. - Use this for classification, summarization, extraction, or any task where you - already have the content and just need LLM reasoning. - - ## Pre-loaded Documents Variable - - If documents were pre-loaded for this session, a `documents` variable is available: - ```python - # documents is a list of dicts with keys: id, title, uri, content - for doc in documents: - print(doc['title'], len(doc['content'])) - ``` - Check if it exists with: `if 'documents' in dir(): ...` - - ## Standard Library Modules - You can import any Python standard library module. - - ## Strategy Guide - - 1. **Explore First**: Start by listing documents or searching to understand what's available. Document names may differ from filenames (e.g., "tbmed593.pdf" might be stored as "TB MED 593" or similar). - 2. **If get_document returns None**: Use `list_documents()` to see actual document titles, or `search()` to find relevant content. - 3. **Iterative Refinement**: Run code, examine results, adjust your approach based on what you find. - 4. **Use print() Liberally**: The REPL captures stdout - print intermediate results to see what you're working with. - 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and use collections. - 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use llm(). - 7. **Cite Your Sources**: Track which documents/chunks informed your answer for citation. - - ## DoclingDocument API - - When you call `get_docling_document(id_or_title)`, you get a DoclingDocument object for structured document analysis. - - ### Properties - - `doc.texts` - List of all text items (paragraphs, headings, etc.) - - `doc.tables` - List of all tables - - `doc.pictures` - List of all pictures/figures - - `doc.name` - Document name - - ### Methods - - `doc.iterate_items(with_groups=False)` - Iterate all items with hierarchy level - Returns tuples of (item, level) where level is nesting depth - - `doc.export_to_markdown()` - Export entire document as markdown string - - ### Text Item Properties - - `item.text` - The text content - - `item.label` - Type: title, paragraph, section_header, list_item, etc. (lowercase enum values) - - `item.prov` - Provenance (page numbers, bounding boxes) - - ### Table Access - - `table.data.num_rows`, `table.data.num_cols` - Dimensions - - `table.data.table_cells` - List of TableCell objects - - `cell.text`, `cell.start_row_offset_idx`, `cell.start_col_offset_idx` - - ### Example Usage - ```python - doc = get_docling_document("My Document") - - # Get all headings - headings = [t.text for t in doc.texts if "header" in str(t.label)] - - # Iterate with structure - for item, level in doc.iterate_items(): - print(" " * level + item.text[:50]) - - # Extract table data - for table in doc.tables: - for cell in table.data.table_cells: - print(f"Row {cell.start_row_offset_idx}, Col {cell.start_col_offset_idx}: {cell.text}") - ``` - - ## Example Patterns - - ### Counting documents matching a condition - ```python - docs = list_documents(limit=100) - count = 0 - for doc in docs: - content = get_document(doc['id']) - if content and 'keyword' in content.lower(): - count += 1 - print(f"Found in: {doc['title']}") - print(f"Total: {count}") - ``` - - ### Aggregating data across documents - ```python - import re - numbers = [] - results = search("financial data", limit=20) - for r in results: - matches = re.findall(r'\$([\d,]+)', r['content']) - for m in matches: - numbers.append(int(m.replace(',', ''))) - print(f"Average: ${sum(numbers)/len(numbers):,.2f}") - ``` - - ### Using llm() for classification - ```python - # Get document content - content = get_document("Q1 Report") - # Use llm() to classify sentiment - sentiment = llm(f"Classify the sentiment as positive, negative, or mixed: {content}") - print(sentiment) - ``` - - ## Workflow - - 1. **ALWAYS start by using execute_code** to explore the knowledge base - 2. Run multiple code blocks as needed to gather information - 3. After collecting data, provide your final answer - - ## Output Format - - CRITICAL: Your final response MUST be valid JSON matching this exact schema: - ```json - {"answer": "Your complete answer here as a string", "program": "Your final consolidated program here as a string"} - ``` - - - `answer`: A clear answer to the user's question with key findings and references to specific documents/chunks. - - `program`: A single, self-contained Python program that produces the answer. Consolidate your exploratory code executions into one clean script. - - Do NOT return arbitrary JSON structures. Always use the exact format: {"answer": "...", "program": "..."} - - CRITICAL: You MUST call execute_code at least once before providing your answer. Never give up without trying to execute code first. - role: system - - content: Search for content about document element types or labels. What are all the different document element types - mentioned? List them all. - role: user - model: gpt-oss - reasoning_effort: low - stream: false - tool_choice: auto - tools: - - function: - description: |- - Execute Python code in a Docker-sandboxed environment. - - The code has access to haiku.rag functions (search, list_documents, - get_document, get_docling_document, llm) and any Python standard - library module. - - Use print() to output results. - - Structured result with success status, stdout, and stderr. - - name: execute_code - parameters: - additionalProperties: false - properties: - code: - description: Python code to execute. - type: string - required: - - code - type: object - strict: true - type: function - - function: - description: Result from RLM agent execution. - name: final_result - parameters: - additionalProperties: false - properties: - answer: - description: The answer to the user's question - type: string - program: - description: The final consolidated program - type: string - required: - - answer - - program - type: object - strict: true - type: function - uri: http://localhost:11434/v1/chat/completions - response: - headers: - content-length: - - '729' + - '683' content-type: - application/json parsed_body: @@ -647,25 +379,25 @@ interactions: index: 0 message: content: '' - reasoning: We need search for "document element types" or "labels". We'll run search. + reasoning: Need to search for document element types or labels. Use search. role: assistant tool_calls: - function: - arguments: '{"code":"results = search(''document element types labels'', limit=20)\nprint(len(results))\nfor - r in results:\n print(r[''document_title''], r[''chunk_id''], r[''score''])\n print(r[''content''][:200])\n print(''---'')"}' + arguments: '{"code":"results = await search(\"document element types\" , limit=20)\nprint(len(results))\nfor + r in results[:5]:\n print(r[''chunk_id''], r[''document_title''], r[''content''][:200])\n"}' name: execute_code - id: call_o2vpon36 + id: call_j7uw3i6n index: 0 type: function - created: 1770373425 - id: chatcmpl-613 + created: 1771924566 + id: chatcmpl-52 model: gpt-oss object: chat.completion system_fingerprint: fp_ollama usage: - completion_tokens: 104 - prompt_tokens: 1763 - total_tokens: 1867 + completion_tokens: 87 + prompt_tokens: 1639 + total_tokens: 1726 status: code: 200 message: OK @@ -678,7 +410,47 @@ interactions: connection: - keep-alive content-length: - - '13324' + - '92' + content-type: + - application/json + host: + - localhost:11434 + method: POST + parsed_body: + encoding_format: base64 + input: + - document element types + model: qwen3-embedding:4b + uri: http://localhost:11434/v1/embeddings + response: + headers: + content-type: + - application/json + transfer-encoding: + - chunked + parsed_body: + data: + - embedding: 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 + index: 0 + object: embedding + model: qwen3-embedding:4b + object: list + usage: + prompt_tokens: 4 + total_tokens: 4 + status: + code: 200 + message: OK +- request: + headers: + accept: + - application/json + accept-encoding: + - gzip, deflate, zstd + connection: + - keep-alive + content-length: + - '9384' content-type: - application/json host: @@ -691,32 +463,33 @@ interactions: IMPORTANT: You MUST use the `execute_code` tool to run Python code. The functions described below are ONLY available inside the execute_code tool - you cannot access them any other way. Always execute code to answer questions; do not just describe what code would do. - CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just use them directly: - - search("query") ✓ CORRECT + CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just call them with `await`: + - results = await search("query") ✓ CORRECT - from haiku.rag import search ✗ WRONG - will fail + - results = search("query") ✗ WRONG - must use await - You have access to a sandboxed Python environment with these haiku.rag functions (use them directly, no imports needed): + You have access to a sandboxed Python interpreter with these haiku.rag functions (use them directly with `await`, no imports needed): ## Available Functions - ### search(query, limit=10) -> list[dict] + ### await search(query, limit=10) -> list[dict] Search the knowledge base using hybrid search (vector + full-text). Returns list of dicts with keys: chunk_id, content, document_id, document_title, document_uri, score, page_numbers, headings - ### list_documents(limit=10, offset=0) -> list[dict] + ### await list_documents(limit=10, offset=0) -> list[dict] List available documents in the knowledge base. Returns list of dicts with keys: id, title, uri, created_at - ### get_document(id_or_title) -> str | None + ### await get_document(id_or_title) -> str | None Get the full text content of a document by ID, title, or URI. Returns the document content as a string, or None if not found. - ### get_docling_document(id_or_title) -> DoclingDocument | None - Get the structured DoclingDocument object for advanced analysis. - Returns a DoclingDocument object, or None if not found. - See "DoclingDocument API" section below for how to use it. + ### await get_chunk(chunk_id) -> dict | None + Get a specific chunk by its ID (from search results). + Returns dict with keys: chunk_id, content, document_id, document_title, headings, page_numbers, labels + Use this to retrieve full chunk details and metadata for citation. - ### llm(prompt) -> str + ### await llm(prompt) -> str Call an LLM directly with the given prompt. Returns the response as a string. Use this for classification, summarization, extraction, or any task where you already have the content and just need LLM reasoning. @@ -731,93 +504,64 @@ interactions: ``` Check if it exists with: `if 'documents' in dir(): ...` - ## Standard Library Modules - You can import any Python standard library module. + ## Available Python Features + + The interpreter supports: variables, arithmetic, strings, f-strings, lists, dicts, tuples, sets, loops, conditionals, comprehensions, functions, async/await, `map()`, `sorted()`/`.sort(key=...)`, try/except, and the `json` module. + + Not supported: imports (other than `json`), class definitions, generators/yield, match statements, decorators, `with` statements. + + For pattern matching or text extraction, use string methods (`str.split`, `str.find`, `str.startswith`, `in` operator) or the `llm()` function. ## Strategy Guide 1. **Explore First**: Start by listing documents or searching to understand what's available. Document names may differ from filenames (e.g., "tbmed593.pdf" might be stored as "TB MED 593" or similar). - 2. **If get_document returns None**: Use `list_documents()` to see actual document titles, or `search()` to find relevant content. + 2. **If get_document returns None**: Use `await list_documents()` to see actual document titles, or `await search()` to find relevant content. 3. **Iterative Refinement**: Run code, examine results, adjust your approach based on what you find. - 4. **Use print() Liberally**: The REPL captures stdout - print intermediate results to see what you're working with. - 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and use collections. - 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use llm(). - 7. **Cite Your Sources**: Track which documents/chunks informed your answer for citation. - - ## DoclingDocument API - - When you call `get_docling_document(id_or_title)`, you get a DoclingDocument object for structured document analysis. - - ### Properties - - `doc.texts` - List of all text items (paragraphs, headings, etc.) - - `doc.tables` - List of all tables - - `doc.pictures` - List of all pictures/figures - - `doc.name` - Document name - - ### Methods - - `doc.iterate_items(with_groups=False)` - Iterate all items with hierarchy level - Returns tuples of (item, level) where level is nesting depth - - `doc.export_to_markdown()` - Export entire document as markdown string - - ### Text Item Properties - - `item.text` - The text content - - `item.label` - Type: title, paragraph, section_header, list_item, etc. (lowercase enum values) - - `item.prov` - Provenance (page numbers, bounding boxes) - - ### Table Access - - `table.data.num_rows`, `table.data.num_cols` - Dimensions - - `table.data.table_cells` - List of TableCell objects - - `cell.text`, `cell.start_row_offset_idx`, `cell.start_col_offset_idx` - - ### Example Usage - ```python - doc = get_docling_document("My Document") - - # Get all headings - headings = [t.text for t in doc.texts if "header" in str(t.label)] - - # Iterate with structure - for item, level in doc.iterate_items(): - print(" " * level + item.text[:50]) - - # Extract table data - for table in doc.tables: - for cell in table.data.table_cells: - print(f"Row {cell.start_row_offset_idx}, Col {cell.start_col_offset_idx}: {cell.text}") - ``` + 4. **Use print() Liberally**: The sandbox captures stdout - print intermediate results to see what you're working with. + 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and data structures. + 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use `await llm()`. + 7. **Cite Your Sources**: Use get_chunk() to retrieve chunk metadata for citations. Track which documents/chunks informed your answer. ## Example Patterns ### Counting documents matching a condition ```python - docs = list_documents(limit=100) + docs = await list_documents(limit=100) count = 0 for doc in docs: - content = get_document(doc['id']) + content = await get_document(doc['id']) if content and 'keyword' in content.lower(): count += 1 print(f"Found in: {doc['title']}") print(f"Total: {count}") ``` - ### Aggregating data across documents + ### Extracting data with llm() ```python - import re numbers = [] - results = search("financial data", limit=20) + results = await search("financial data", limit=20) for r in results: - matches = re.findall(r'\$([\d,]+)', r['content']) - for m in matches: - numbers.append(int(m.replace(',', ''))) - print(f"Average: ${sum(numbers)/len(numbers):,.2f}") + extracted = await llm(f"Extract all dollar amounts from this text as a comma-separated list of numbers (no $ signs): {r['content']}") + for part in extracted.split(','): + part = part.strip().replace(',', '') + if part.isdigit(): + numbers.append(int(part)) + if numbers: + print(f"Average: {sum(numbers) / len(numbers)}") + ``` + + ### Using search results with get_chunk for citations + ```python + results = await search("safety requirements", limit=5) + for r in results: + chunk = await get_chunk(r['chunk_id']) + print(f"From '{chunk['document_title']}', page {chunk['page_numbers']}: {chunk['content'][:100]}") ``` ### Using llm() for classification ```python - # Get document content - content = get_document("Q1 Report") - # Use llm() to classify sentiment - sentiment = llm(f"Classify the sentiment as positive, negative, or mixed: {content}") + content = await get_document("Q1 Report") + sentiment = await llm(f"Classify the sentiment as positive, negative, or mixed: {content}") print(sentiment) ``` @@ -844,61 +588,30 @@ interactions: - content: Search for content about document element types or labels. What are all the different document element types mentioned? List them all. role: user - - content: |- - - We need search for "document element types" or "labels". We'll run search. - + - content: null + reasoning: Need to search for document element types or labels. Use search. role: assistant tool_calls: - function: - arguments: '{"code":"results = search(''document element types labels'', limit=20)\nprint(len(results))\nfor r - in results:\n print(r[''document_title''], r[''chunk_id''], r[''score''])\n print(r[''content''][:200])\n print(''---'')"}' + arguments: '{"code":"results = await search(\"document element types\" , limit=20)\nprint(len(results))\nfor r + in results[:5]:\n print(r[''chunk_id''], r[''document_title''], r[''content''][:200])\n"}' name: execute_code - id: call_o2vpon36 + id: call_j7uw3i6n type: function - - content: '{"code":"results = search(''document element types labels'', limit=20)\nprint(len(results))\nfor r in results:\n print(r[''document_title''], - r[''chunk_id''], r[''score''])\n print(r[''content''][:200])\n print(''---'')","stdout":"17\nNone 360a55ef-1fae-45c7-9506-d681bf7d2642 - 0.032786883413791656\nPhase 2: Label selection and guideline. We reviewed the collected documents and identified - the most common structural features they exhibit. This was achieved by identifying recurrent layout elements \n---\nNone - f71a4730-2d9c-4317-9e33-2488adda8f3d 0.0317540317773819\n\nTable 1: DocLayNet dataset overview. Along with the frequency - of each class label, we present the relative occurrence (as % of row \"Total\") in the train, test and validation - sets. The inter-annotator\n---\nNone 994f8aeb-bdf3-434d-9b2e-69d4a6a9c623 0.03015873022377491\n$_{Affiliation}$, - as seen in DocBank, are often only distinguishable by discriminating on\nPreparation work included uploading and - parsing the sourced PDF documents in the Corpus Conversion Service (CC\n---\nNone 70a1c951-bc93-4302-95ca-bdbb832f3cf9 - 0.029462365433573723\nPhase 1: Data selection and preparation. Our inclusion criteria for documents were described - in Section 3. A large effort went into ensuring that all documents are free to use. The data sources includ\n---\nNone - 916ed8c5-d868-4064-a459-1f2cc704df4e 0.028371628373861313\nmAP @ 0.5-0.95 (%).Sci = 89-94. Total, triple inter-annotator - mAP @ 0.5-0.95 (%).Law = 86-91. Total, triple inter-annotator mAP @ 0.5-0.95 (%).Pat = 71-76. Total, triple inter-annotator - mAP @ 0.5-0.95\n---\nNone 41a5b3f5-ff96-4856-9eb9-4695fe28b39c 0.01587301678955555\nCaption, Count = 22524. Caption, - % of Total.Train = 2.04. Caption, % of Total.Test = 1.77. Caption, % of Total.Val = 2.32. Caption, triple inter-annotator - mAP @ 0.5-0.95 (%).All = 84-89. Caption, trip\n---\nNone c359d67a-0809-45bb-bfb0-139817b967fd 0.015625\nPage-footer, - triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 96-98. Page-header, Count = 58022. Page-header, % of Total.Train - = 5.10. Page-header, % of Total.Test = 6.70. Page-header, % of Total.Val =\n---\nNone 50e95bc9-862e-4bdb-9e8c-4d1e38d0eee6 - 0.015384615398943424\n0.5-0.95 (%).Law = 87-94. Section-header, triple inter-annotator mAP @ 0.5-0.95 (%).Pat = - 69-73. Section-header, triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 78-86. Table, Count = 34733. Table, % o\n---\nNone - d40d3add-cd91-4774-9d3e-77c388e0f9a4 0.01515151560306549\n89-93. Text, triple inter-annotator mAP @ 0.5-0.95 (%).Law - = 87-92. Text, triple inter-annotator mAP @ 0.5-0.95 (%).Pat = 71-79. Text, triple inter-annotator mAP @ 0.5-0.95 - (%).Ten = 87-95. Title, Cou\n---\nNone 683d9ab8-6363-4165-9cc7-b89145db3f33 0.014925372786819935\n185660. List-item, - % of Total.Train = 17.19. List-item, % of Total.Test = 13.34. List-item, % of Total.Val = 15.82. List-item, triple - inter-annotator mAP @ 0.5-0.95 (%).All = 87-88. List-item, triple \n---\nNone f5d1d638-002e-42ab-8c35-6fbf834ab435 - 0.014705882407724857\ninter-annotator mAP @ 0.5-0.95 (%).Pat = 66-71. Picture, triple inter-annotator mAP @ 0.5-0.95 - (%).Ten = 59-76. Section-header, Count = 142884. Section-header, % of Total.Train = 12.60. Section-header\n---\nNone - 171eb4b0-e518-4e65-9ef4-5655789dceae 0.014492753893136978\nn/a. Footnote, Count = 6318. Footnote, % of Total.Train - = 0.60. Footnote, % of Total.Test = 0.31. Footnote, % of Total.Val = 0.58. Footnote, triple inter-annotator mAP - @ 0.5-0.95 (%).All = 83-91. Foot\n---\nNone 64741ed7-d686-4e6c-b8e1-462a51282ecc 0.014084506779909134\n(%).Sci = - 94-100. Title, triple inter-annotator mAP @ 0.5-0.95 (%).Law = 82-96. Title, triple inter-annotator mAP @ 0.5-0.95 - (%).Pat = 68-79. Title, triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 24-\n---\nNone d8771358-ce2e-407f-af30-ee410ed8a579 - 0.013888888992369175\n0.5-0.95 (%).Pat = 97-99. Page-header, triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 81-86. - Picture, Count = 45976. Picture, % of Total.Train = 4.21. Picture, % of Total.Test = 2.78. Picture, % of T\n---\nNone - c2289d48-99bf-4400-ac39-155b764485db 0.013698630034923553\ninter-annotator mAP @ 0.5-0.95 (%).Law = 58-80. Table, - triple inter-annotator mAP @ 0.5-0.95 (%).Pat = 79-84. Table, triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 70-85. - Text, Count = 510377. Text, \n---\nNone 6a21d905-7202-4ed9-9241-32934c09f36e 0.013513513840734959\n0.5-0.95 (%).Ten - = 93-95. Page-footer, Count = 70878. Page-footer, % of Total.Train = 6.51. Page-footer, % of Total.Test = 5.58. - Page-footer, % of Total.Val = 6.00. Page-footer, triple inter-annotator\n---\nNone 610e94fb-1051-4e09-b94b-98cd485e551f - 0.01315789483487606\n= 82-97. Formula, Count = 25027. Formula, % of Total.Train = 2.25. Formula, % of Total.Test - = 1.90. Formula, % of Total.Val = 2.96. Formula, triple inter-annotator mAP @ 0.5-0.95 (%).All = 83-85. Form\n---\n","stderr":"","success":true}' + - content: '{"code":"results = await search(\"document element types\" , limit=20)\nprint(len(results))\nfor r in results[:5]:\n print(r[''chunk_id''], + r[''document_title''], r[''content''][:200])\n","stdout":"17\n7c3a912c-a769-49e9-a661-402d7d64ec9f None Phase 2: + Label selection and guideline. We reviewed the collected documents and identified the most common structural features + they exhibit. This was achieved by identifying recurrent layout elements \n62e98936-a694-4cb0-a1a4-409b0121e08d + None Phase 1: Data selection and preparation. Our inclusion criteria for documents were described in Section 3. + A large effort went into ensuring that all documents are free to use. The data sources includ\nd583f0ed-c230-4f8e-988a-86875c0c633d + None $_{Affiliation}$, as seen in DocBank, are often only distinguishable by discriminating on\nPreparation work + included uploading and parsing the sourced PDF documents in the Corpus Conversion Service (CC\n180df00a-2a5b-4e2a-bfcc-4cd67fbc9db0 + None \nTable 1: DocLayNet dataset overview. Along with the frequency of each class label, we present the relative + occurrence (as % of row \"Total\") in the train, test and validation sets. The inter-annotator\nf3c18411-85df-4c72-a784-f39f602745ec + None Page-footer, triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 96-98. Page-header, Count = 58022. Page-header, + % of Total.Train = 5.10. Page-header, % of Total.Test = 6.70. Page-header, % of Total.Val =\n","stderr":"","success":true}' role: tool - tool_call_id: call_o2vpon36 + tool_call_id: call_j7uw3i6n model: gpt-oss reasoning_effort: low stream: false @@ -906,11 +619,10 @@ interactions: tools: - function: description: |- - Execute Python code in a Docker-sandboxed environment. + Execute Python code in a sandboxed interpreter. The code has access to haiku.rag functions (search, list_documents, - get_document, get_docling_document, llm) and any Python standard - library module. + get_document, get_chunk, llm). Use print() to output results. @@ -950,7 +662,7 @@ interactions: response: headers: content-length: - - '860' + - '891' content-type: - application/json parsed_body: @@ -959,26 +671,27 @@ interactions: index: 0 message: content: '' - reasoning: Chapters mention labels like "page-header", "page-footer", "section-header", "picture", "table", "text", - "title", "list-item", "footnote", "formula". Also maybe "Caption". We'll extract unique set. + reasoning: We need list all document element types or labels mentioned. Likely these are labels in documents like + "Page-header", "Page-footer", etc. Let's search more. role: assistant tool_calls: - function: - arguments: '{"code":"labels=set()\nfor r in results:\n if ''label'' in r[''content'']:\n pass\n","stdout":"","stderr":"FileNotFoundError: - [Errno 2] No such file or directory: ''results''\n","success":false}' + arguments: '{"code":"results = await search(\"label\" , limit=30)\nlabels=set()\nfor r in results:\n # look + for uppercase words separated by hyphens\n import re\n for m in re.findall(r\"([A-Z][A-Za-z]+(?:-[A-Z][A-Za-z]+)*)\", + r[''content'']):\n labels.add(m)\nprint(sorted(labels))\n"}' name: execute_code - id: call_1mr5hrfc + id: call_gzg850pl index: 0 type: function - created: 1770373431 - id: chatcmpl-908 + created: 1771924570 + id: chatcmpl-449 model: gpt-oss object: chat.completion system_fingerprint: fp_ollama usage: - completion_tokens: 128 - prompt_tokens: 3814 - total_tokens: 3942 + completion_tokens: 142 + prompt_tokens: 2158 + total_tokens: 2300 status: code: 200 message: OK @@ -991,7 +704,47 @@ interactions: connection: - keep-alive content-length: - - '14601' + - '75' + content-type: + - application/json + host: + - localhost:11434 + method: POST + parsed_body: + encoding_format: base64 + input: + - label + model: qwen3-embedding:4b + uri: http://localhost:11434/v1/embeddings + response: + headers: + content-type: + - application/json + transfer-encoding: + - chunked + parsed_body: + data: + - embedding: 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 + index: 0 + object: embedding + model: qwen3-embedding:4b + object: list + usage: + prompt_tokens: 2 + total_tokens: 2 + status: + code: 200 + message: OK +- request: + headers: + accept: + - application/json + accept-encoding: + - gzip, deflate, zstd + connection: + - keep-alive + content-length: + - '10466' content-type: - application/json host: @@ -1004,32 +757,33 @@ interactions: IMPORTANT: You MUST use the `execute_code` tool to run Python code. The functions described below are ONLY available inside the execute_code tool - you cannot access them any other way. Always execute code to answer questions; do not just describe what code would do. - CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just use them directly: - - search("query") ✓ CORRECT + CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just call them with `await`: + - results = await search("query") ✓ CORRECT - from haiku.rag import search ✗ WRONG - will fail + - results = search("query") ✗ WRONG - must use await - You have access to a sandboxed Python environment with these haiku.rag functions (use them directly, no imports needed): + You have access to a sandboxed Python interpreter with these haiku.rag functions (use them directly with `await`, no imports needed): ## Available Functions - ### search(query, limit=10) -> list[dict] + ### await search(query, limit=10) -> list[dict] Search the knowledge base using hybrid search (vector + full-text). Returns list of dicts with keys: chunk_id, content, document_id, document_title, document_uri, score, page_numbers, headings - ### list_documents(limit=10, offset=0) -> list[dict] + ### await list_documents(limit=10, offset=0) -> list[dict] List available documents in the knowledge base. Returns list of dicts with keys: id, title, uri, created_at - ### get_document(id_or_title) -> str | None + ### await get_document(id_or_title) -> str | None Get the full text content of a document by ID, title, or URI. Returns the document content as a string, or None if not found. - ### get_docling_document(id_or_title) -> DoclingDocument | None - Get the structured DoclingDocument object for advanced analysis. - Returns a DoclingDocument object, or None if not found. - See "DoclingDocument API" section below for how to use it. + ### await get_chunk(chunk_id) -> dict | None + Get a specific chunk by its ID (from search results). + Returns dict with keys: chunk_id, content, document_id, document_title, headings, page_numbers, labels + Use this to retrieve full chunk details and metadata for citation. - ### llm(prompt) -> str + ### await llm(prompt) -> str Call an LLM directly with the given prompt. Returns the response as a string. Use this for classification, summarization, extraction, or any task where you already have the content and just need LLM reasoning. @@ -1044,93 +798,64 @@ interactions: ``` Check if it exists with: `if 'documents' in dir(): ...` - ## Standard Library Modules - You can import any Python standard library module. + ## Available Python Features + + The interpreter supports: variables, arithmetic, strings, f-strings, lists, dicts, tuples, sets, loops, conditionals, comprehensions, functions, async/await, `map()`, `sorted()`/`.sort(key=...)`, try/except, and the `json` module. + + Not supported: imports (other than `json`), class definitions, generators/yield, match statements, decorators, `with` statements. + + For pattern matching or text extraction, use string methods (`str.split`, `str.find`, `str.startswith`, `in` operator) or the `llm()` function. ## Strategy Guide 1. **Explore First**: Start by listing documents or searching to understand what's available. Document names may differ from filenames (e.g., "tbmed593.pdf" might be stored as "TB MED 593" or similar). - 2. **If get_document returns None**: Use `list_documents()` to see actual document titles, or `search()` to find relevant content. + 2. **If get_document returns None**: Use `await list_documents()` to see actual document titles, or `await search()` to find relevant content. 3. **Iterative Refinement**: Run code, examine results, adjust your approach based on what you find. - 4. **Use print() Liberally**: The REPL captures stdout - print intermediate results to see what you're working with. - 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and use collections. - 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use llm(). - 7. **Cite Your Sources**: Track which documents/chunks informed your answer for citation. - - ## DoclingDocument API - - When you call `get_docling_document(id_or_title)`, you get a DoclingDocument object for structured document analysis. - - ### Properties - - `doc.texts` - List of all text items (paragraphs, headings, etc.) - - `doc.tables` - List of all tables - - `doc.pictures` - List of all pictures/figures - - `doc.name` - Document name - - ### Methods - - `doc.iterate_items(with_groups=False)` - Iterate all items with hierarchy level - Returns tuples of (item, level) where level is nesting depth - - `doc.export_to_markdown()` - Export entire document as markdown string - - ### Text Item Properties - - `item.text` - The text content - - `item.label` - Type: title, paragraph, section_header, list_item, etc. (lowercase enum values) - - `item.prov` - Provenance (page numbers, bounding boxes) - - ### Table Access - - `table.data.num_rows`, `table.data.num_cols` - Dimensions - - `table.data.table_cells` - List of TableCell objects - - `cell.text`, `cell.start_row_offset_idx`, `cell.start_col_offset_idx` - - ### Example Usage - ```python - doc = get_docling_document("My Document") - - # Get all headings - headings = [t.text for t in doc.texts if "header" in str(t.label)] - - # Iterate with structure - for item, level in doc.iterate_items(): - print(" " * level + item.text[:50]) - - # Extract table data - for table in doc.tables: - for cell in table.data.table_cells: - print(f"Row {cell.start_row_offset_idx}, Col {cell.start_col_offset_idx}: {cell.text}") - ``` + 4. **Use print() Liberally**: The sandbox captures stdout - print intermediate results to see what you're working with. + 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and data structures. + 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use `await llm()`. + 7. **Cite Your Sources**: Use get_chunk() to retrieve chunk metadata for citations. Track which documents/chunks informed your answer. ## Example Patterns ### Counting documents matching a condition ```python - docs = list_documents(limit=100) + docs = await list_documents(limit=100) count = 0 for doc in docs: - content = get_document(doc['id']) + content = await get_document(doc['id']) if content and 'keyword' in content.lower(): count += 1 print(f"Found in: {doc['title']}") print(f"Total: {count}") ``` - ### Aggregating data across documents + ### Extracting data with llm() ```python - import re numbers = [] - results = search("financial data", limit=20) + results = await search("financial data", limit=20) for r in results: - matches = re.findall(r'\$([\d,]+)', r['content']) - for m in matches: - numbers.append(int(m.replace(',', ''))) - print(f"Average: ${sum(numbers)/len(numbers):,.2f}") + extracted = await llm(f"Extract all dollar amounts from this text as a comma-separated list of numbers (no $ signs): {r['content']}") + for part in extracted.split(','): + part = part.strip().replace(',', '') + if part.isdigit(): + numbers.append(int(part)) + if numbers: + print(f"Average: {sum(numbers) / len(numbers)}") + ``` + + ### Using search results with get_chunk for citations + ```python + results = await search("safety requirements", limit=5) + for r in results: + chunk = await get_chunk(r['chunk_id']) + print(f"From '{chunk['document_title']}', page {chunk['page_numbers']}: {chunk['content'][:100]}") ``` ### Using llm() for classification ```python - # Get document content - content = get_document("Q1 Report") - # Use llm() to classify sentiment - sentiment = llm(f"Classify the sentiment as positive, negative, or mixed: {content}") + content = await get_document("Q1 Report") + sentiment = await llm(f"Classify the sentiment as positive, negative, or mixed: {content}") print(sentiment) ``` @@ -1157,107 +882,47 @@ interactions: - content: Search for content about document element types or labels. What are all the different document element types mentioned? List them all. role: user - - content: |- - - We need search for "document element types" or "labels". We'll run search. - + - content: null + reasoning: Need to search for document element types or labels. Use search. role: assistant tool_calls: - function: - arguments: '{"code":"results = search(''document element types labels'', limit=20)\nprint(len(results))\nfor r - in results:\n print(r[''document_title''], r[''chunk_id''], r[''score''])\n print(r[''content''][:200])\n print(''---'')"}' + arguments: '{"code":"results = await search(\"document element types\" , limit=20)\nprint(len(results))\nfor r + in results[:5]:\n print(r[''chunk_id''], r[''document_title''], r[''content''][:200])\n"}' name: execute_code - id: call_o2vpon36 + id: call_j7uw3i6n type: function - - content: '{"code":"results = search(''document element types labels'', limit=20)\nprint(len(results))\nfor r in results:\n print(r[''document_title''], - r[''chunk_id''], r[''score''])\n print(r[''content''][:200])\n print(''---'')","stdout":"17\nNone 360a55ef-1fae-45c7-9506-d681bf7d2642 - 0.032786883413791656\nPhase 2: Label selection and guideline. We reviewed the collected documents and identified - the most common structural features they exhibit. This was achieved by identifying recurrent layout elements \n---\nNone - f71a4730-2d9c-4317-9e33-2488adda8f3d 0.0317540317773819\n\nTable 1: DocLayNet dataset overview. Along with the frequency - of each class label, we present the relative occurrence (as % of row \"Total\") in the train, test and validation - sets. The inter-annotator\n---\nNone 994f8aeb-bdf3-434d-9b2e-69d4a6a9c623 0.03015873022377491\n$_{Affiliation}$, - as seen in DocBank, are often only distinguishable by discriminating on\nPreparation work included uploading and - parsing the sourced PDF documents in the Corpus Conversion Service (CC\n---\nNone 70a1c951-bc93-4302-95ca-bdbb832f3cf9 - 0.029462365433573723\nPhase 1: Data selection and preparation. Our inclusion criteria for documents were described - in Section 3. A large effort went into ensuring that all documents are free to use. The data sources includ\n---\nNone - 916ed8c5-d868-4064-a459-1f2cc704df4e 0.028371628373861313\nmAP @ 0.5-0.95 (%).Sci = 89-94. Total, triple inter-annotator - mAP @ 0.5-0.95 (%).Law = 86-91. Total, triple inter-annotator mAP @ 0.5-0.95 (%).Pat = 71-76. Total, triple inter-annotator - mAP @ 0.5-0.95\n---\nNone 41a5b3f5-ff96-4856-9eb9-4695fe28b39c 0.01587301678955555\nCaption, Count = 22524. Caption, - % of Total.Train = 2.04. Caption, % of Total.Test = 1.77. Caption, % of Total.Val = 2.32. Caption, triple inter-annotator - mAP @ 0.5-0.95 (%).All = 84-89. Caption, trip\n---\nNone c359d67a-0809-45bb-bfb0-139817b967fd 0.015625\nPage-footer, - triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 96-98. Page-header, Count = 58022. Page-header, % of Total.Train - = 5.10. Page-header, % of Total.Test = 6.70. Page-header, % of Total.Val =\n---\nNone 50e95bc9-862e-4bdb-9e8c-4d1e38d0eee6 - 0.015384615398943424\n0.5-0.95 (%).Law = 87-94. Section-header, triple inter-annotator mAP @ 0.5-0.95 (%).Pat = - 69-73. Section-header, triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 78-86. Table, Count = 34733. Table, % o\n---\nNone - d40d3add-cd91-4774-9d3e-77c388e0f9a4 0.01515151560306549\n89-93. Text, triple inter-annotator mAP @ 0.5-0.95 (%).Law - = 87-92. Text, triple inter-annotator mAP @ 0.5-0.95 (%).Pat = 71-79. Text, triple inter-annotator mAP @ 0.5-0.95 - (%).Ten = 87-95. Title, Cou\n---\nNone 683d9ab8-6363-4165-9cc7-b89145db3f33 0.014925372786819935\n185660. List-item, - % of Total.Train = 17.19. List-item, % of Total.Test = 13.34. List-item, % of Total.Val = 15.82. List-item, triple - inter-annotator mAP @ 0.5-0.95 (%).All = 87-88. List-item, triple \n---\nNone f5d1d638-002e-42ab-8c35-6fbf834ab435 - 0.014705882407724857\ninter-annotator mAP @ 0.5-0.95 (%).Pat = 66-71. Picture, triple inter-annotator mAP @ 0.5-0.95 - (%).Ten = 59-76. Section-header, Count = 142884. Section-header, % of Total.Train = 12.60. Section-header\n---\nNone - 171eb4b0-e518-4e65-9ef4-5655789dceae 0.014492753893136978\nn/a. Footnote, Count = 6318. Footnote, % of Total.Train - = 0.60. Footnote, % of Total.Test = 0.31. Footnote, % of Total.Val = 0.58. Footnote, triple inter-annotator mAP - @ 0.5-0.95 (%).All = 83-91. Foot\n---\nNone 64741ed7-d686-4e6c-b8e1-462a51282ecc 0.014084506779909134\n(%).Sci = - 94-100. Title, triple inter-annotator mAP @ 0.5-0.95 (%).Law = 82-96. Title, triple inter-annotator mAP @ 0.5-0.95 - (%).Pat = 68-79. Title, triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 24-\n---\nNone d8771358-ce2e-407f-af30-ee410ed8a579 - 0.013888888992369175\n0.5-0.95 (%).Pat = 97-99. Page-header, triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 81-86. - Picture, Count = 45976. Picture, % of Total.Train = 4.21. Picture, % of Total.Test = 2.78. Picture, % of T\n---\nNone - c2289d48-99bf-4400-ac39-155b764485db 0.013698630034923553\ninter-annotator mAP @ 0.5-0.95 (%).Law = 58-80. Table, - triple inter-annotator mAP @ 0.5-0.95 (%).Pat = 79-84. Table, triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 70-85. - Text, Count = 510377. Text, \n---\nNone 6a21d905-7202-4ed9-9241-32934c09f36e 0.013513513840734959\n0.5-0.95 (%).Ten - = 93-95. Page-footer, Count = 70878. Page-footer, % of Total.Train = 6.51. Page-footer, % of Total.Test = 5.58. - Page-footer, % of Total.Val = 6.00. Page-footer, triple inter-annotator\n---\nNone 610e94fb-1051-4e09-b94b-98cd485e551f - 0.01315789483487606\n= 82-97. Formula, Count = 25027. Formula, % of Total.Train = 2.25. Formula, % of Total.Test - = 1.90. Formula, % of Total.Val = 2.96. Formula, triple inter-annotator mAP @ 0.5-0.95 (%).All = 83-85. Form\n---\n","stderr":"","success":true}' + - content: '{"code":"results = await search(\"document element types\" , limit=20)\nprint(len(results))\nfor r in results[:5]:\n print(r[''chunk_id''], + r[''document_title''], r[''content''][:200])\n","stdout":"17\n7c3a912c-a769-49e9-a661-402d7d64ec9f None Phase 2: + Label selection and guideline. We reviewed the collected documents and identified the most common structural features + they exhibit. This was achieved by identifying recurrent layout elements \n62e98936-a694-4cb0-a1a4-409b0121e08d + None Phase 1: Data selection and preparation. Our inclusion criteria for documents were described in Section 3. + A large effort went into ensuring that all documents are free to use. The data sources includ\nd583f0ed-c230-4f8e-988a-86875c0c633d + None $_{Affiliation}$, as seen in DocBank, are often only distinguishable by discriminating on\nPreparation work + included uploading and parsing the sourced PDF documents in the Corpus Conversion Service (CC\n180df00a-2a5b-4e2a-bfcc-4cd67fbc9db0 + None \nTable 1: DocLayNet dataset overview. Along with the frequency of each class label, we present the relative + occurrence (as % of row \"Total\") in the train, test and validation sets. The inter-annotator\nf3c18411-85df-4c72-a784-f39f602745ec + None Page-footer, triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 96-98. Page-header, Count = 58022. Page-header, + % of Total.Train = 5.10. Page-header, % of Total.Test = 6.70. Page-header, % of Total.Val =\n","stderr":"","success":true}' role: tool - tool_call_id: call_o2vpon36 - - content: |- - - Chapters mention labels like "page-header", "page-footer", "section-header", "picture", "table", "text", "title", "list-item", "footnote", "formula". Also maybe "Caption". We'll extract unique set. - + tool_call_id: call_j7uw3i6n + - content: null + reasoning: We need list all document element types or labels mentioned. Likely these are labels in documents like + "Page-header", "Page-footer", etc. Let's search more. role: assistant tool_calls: - function: - arguments: '{"code":"labels=set()\nfor r in results:\n if ''label'' in r[''content'']:\n pass\n","stdout":"","stderr":"FileNotFoundError: - [Errno 2] No such file or directory: ''results''\n","success":false}' + arguments: '{"code":"results = await search(\"label\" , limit=30)\nlabels=set()\nfor r in results:\n # look + for uppercase words separated by hyphens\n import re\n for m in re.findall(r\"([A-Z][A-Za-z]+(?:-[A-Z][A-Za-z]+)*)\", + r[''content'']):\n labels.add(m)\nprint(sorted(labels))\n"}' name: execute_code - id: call_1mr5hrfc + id: call_gzg850pl type: function - - content: |- - 3 validation errors: - ```json - [ - { - "type": "extra_forbidden", - "loc": [ - "stdout" - ], - "msg": "Extra inputs are not permitted", - "input": "" - }, - { - "type": "extra_forbidden", - "loc": [ - "stderr" - ], - "msg": "Extra inputs are not permitted", - "input": "FileNotFoundError: [Errno 2] No such file or directory: 'results'\n" - }, - { - "type": "extra_forbidden", - "loc": [ - "success" - ], - "msg": "Extra inputs are not permitted", - "input": false - } - ] - ``` - - Fix the errors and try again. + - content: '{"code":"results = await search(\"label\" , limit=30)\nlabels=set()\nfor r in results:\n # look for uppercase + words separated by hyphens\n import re\n for m in re.findall(r\"([A-Z][A-Za-z]+(?:-[A-Z][A-Za-z]+)*)\", r[''content'']):\n labels.add(m)\nprint(sorted(labels))\n","stdout":"","stderr":"ModuleNotFoundError: + No module named ''re''","success":false}' role: tool - tool_call_id: call_1mr5hrfc + tool_call_id: call_gzg850pl model: gpt-oss reasoning_effort: low stream: false @@ -1265,11 +930,10 @@ interactions: tools: - function: description: |- - Execute Python code in a Docker-sandboxed environment. + Execute Python code in a sandboxed interpreter. The code has access to haiku.rag functions (search, list_documents, - get_document, get_docling_document, llm) and any Python standard - library module. + get_document, get_chunk, llm). Use print() to output results. @@ -1309,7 +973,7 @@ interactions: response: headers: content-length: - - '856' + - '704' content-type: - application/json parsed_body: @@ -1318,26 +982,25 @@ interactions: index: 0 message: content: '' - reasoning: Need to store results variable outside. Let's re-run search and store. + reasoning: Import not allowed. Use string methods. Probably easier to manually inspect. Let's list results. role: assistant tool_calls: - function: - arguments: '{"code":"results = search(''document element types labels'', limit=20)\nlabels=set()\nimport re\nfor - r in results:\n # extract words like \"Caption\", \"Formula\", etc by regex\n for m in re.finditer(r\"\\b([A-Za-z-]+)\\b\", - r[''content'']):\n labels.add(m.group(1))\nprint(sorted(labels)[:30])\nprint(''count'', len(labels))"}' + arguments: '{"code":"results = await search(\"label\" , limit=10)\nfor r in results:\n print(''---'')\n print(r[''chunk_id''], + r[''document_title''])\n print(r[''content''][:400])\n"}' name: execute_code - id: call_u9fkvji9 + id: call_4s9httx6 index: 0 type: function - created: 1770373434 - id: chatcmpl-736 + created: 1771924572 + id: chatcmpl-410 model: gpt-oss object: chat.completion system_fingerprint: fp_ollama usage: - completion_tokens: 133 - prompt_tokens: 4120 - total_tokens: 4253 + completion_tokens: 92 + prompt_tokens: 2421 + total_tokens: 2513 status: code: 200 message: OK @@ -1350,373 +1013,34 @@ interactions: connection: - keep-alive content-length: - - '15963' + - '75' content-type: - application/json host: - localhost:11434 method: POST parsed_body: - messages: - - content: |- - You are a Recursive Language Model (RLM) agent that solves complex research questions by writing and executing Python code. - - IMPORTANT: You MUST use the `execute_code` tool to run Python code. The functions described below are ONLY available inside the execute_code tool - you cannot access them any other way. Always execute code to answer questions; do not just describe what code would do. - - CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just use them directly: - - search("query") ✓ CORRECT - - from haiku.rag import search ✗ WRONG - will fail - - You have access to a sandboxed Python environment with these haiku.rag functions (use them directly, no imports needed): - - ## Available Functions - - ### search(query, limit=10) -> list[dict] - Search the knowledge base using hybrid search (vector + full-text). - Returns list of dicts with keys: chunk_id, content, document_id, document_title, document_uri, score, page_numbers, headings - - ### list_documents(limit=10, offset=0) -> list[dict] - List available documents in the knowledge base. - Returns list of dicts with keys: id, title, uri, created_at - - ### get_document(id_or_title) -> str | None - Get the full text content of a document by ID, title, or URI. - Returns the document content as a string, or None if not found. - - ### get_docling_document(id_or_title) -> DoclingDocument | None - Get the structured DoclingDocument object for advanced analysis. - Returns a DoclingDocument object, or None if not found. - See "DoclingDocument API" section below for how to use it. - - ### llm(prompt) -> str - Call an LLM directly with the given prompt. Returns the response as a string. - Use this for classification, summarization, extraction, or any task where you - already have the content and just need LLM reasoning. - - ## Pre-loaded Documents Variable - - If documents were pre-loaded for this session, a `documents` variable is available: - ```python - # documents is a list of dicts with keys: id, title, uri, content - for doc in documents: - print(doc['title'], len(doc['content'])) - ``` - Check if it exists with: `if 'documents' in dir(): ...` - - ## Standard Library Modules - You can import any Python standard library module. - - ## Strategy Guide - - 1. **Explore First**: Start by listing documents or searching to understand what's available. Document names may differ from filenames (e.g., "tbmed593.pdf" might be stored as "TB MED 593" or similar). - 2. **If get_document returns None**: Use `list_documents()` to see actual document titles, or `search()` to find relevant content. - 3. **Iterative Refinement**: Run code, examine results, adjust your approach based on what you find. - 4. **Use print() Liberally**: The REPL captures stdout - print intermediate results to see what you're working with. - 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and use collections. - 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use llm(). - 7. **Cite Your Sources**: Track which documents/chunks informed your answer for citation. - - ## DoclingDocument API - - When you call `get_docling_document(id_or_title)`, you get a DoclingDocument object for structured document analysis. - - ### Properties - - `doc.texts` - List of all text items (paragraphs, headings, etc.) - - `doc.tables` - List of all tables - - `doc.pictures` - List of all pictures/figures - - `doc.name` - Document name - - ### Methods - - `doc.iterate_items(with_groups=False)` - Iterate all items with hierarchy level - Returns tuples of (item, level) where level is nesting depth - - `doc.export_to_markdown()` - Export entire document as markdown string - - ### Text Item Properties - - `item.text` - The text content - - `item.label` - Type: title, paragraph, section_header, list_item, etc. (lowercase enum values) - - `item.prov` - Provenance (page numbers, bounding boxes) - - ### Table Access - - `table.data.num_rows`, `table.data.num_cols` - Dimensions - - `table.data.table_cells` - List of TableCell objects - - `cell.text`, `cell.start_row_offset_idx`, `cell.start_col_offset_idx` - - ### Example Usage - ```python - doc = get_docling_document("My Document") - - # Get all headings - headings = [t.text for t in doc.texts if "header" in str(t.label)] - - # Iterate with structure - for item, level in doc.iterate_items(): - print(" " * level + item.text[:50]) - - # Extract table data - for table in doc.tables: - for cell in table.data.table_cells: - print(f"Row {cell.start_row_offset_idx}, Col {cell.start_col_offset_idx}: {cell.text}") - ``` - - ## Example Patterns - - ### Counting documents matching a condition - ```python - docs = list_documents(limit=100) - count = 0 - for doc in docs: - content = get_document(doc['id']) - if content and 'keyword' in content.lower(): - count += 1 - print(f"Found in: {doc['title']}") - print(f"Total: {count}") - ``` - - ### Aggregating data across documents - ```python - import re - numbers = [] - results = search("financial data", limit=20) - for r in results: - matches = re.findall(r'\$([\d,]+)', r['content']) - for m in matches: - numbers.append(int(m.replace(',', ''))) - print(f"Average: ${sum(numbers)/len(numbers):,.2f}") - ``` - - ### Using llm() for classification - ```python - # Get document content - content = get_document("Q1 Report") - # Use llm() to classify sentiment - sentiment = llm(f"Classify the sentiment as positive, negative, or mixed: {content}") - print(sentiment) - ``` - - ## Workflow - - 1. **ALWAYS start by using execute_code** to explore the knowledge base - 2. Run multiple code blocks as needed to gather information - 3. After collecting data, provide your final answer - - ## Output Format - - CRITICAL: Your final response MUST be valid JSON matching this exact schema: - ```json - {"answer": "Your complete answer here as a string", "program": "Your final consolidated program here as a string"} - ``` - - - `answer`: A clear answer to the user's question with key findings and references to specific documents/chunks. - - `program`: A single, self-contained Python program that produces the answer. Consolidate your exploratory code executions into one clean script. - - Do NOT return arbitrary JSON structures. Always use the exact format: {"answer": "...", "program": "..."} - - CRITICAL: You MUST call execute_code at least once before providing your answer. Never give up without trying to execute code first. - role: system - - content: Search for content about document element types or labels. What are all the different document element types - mentioned? List them all. - role: user - - content: |- - - We need search for "document element types" or "labels". We'll run search. - - role: assistant - tool_calls: - - function: - arguments: '{"code":"results = search(''document element types labels'', limit=20)\nprint(len(results))\nfor r - in results:\n print(r[''document_title''], r[''chunk_id''], r[''score''])\n print(r[''content''][:200])\n print(''---'')"}' - name: execute_code - id: call_o2vpon36 - type: function - - content: '{"code":"results = search(''document element types labels'', limit=20)\nprint(len(results))\nfor r in results:\n print(r[''document_title''], - r[''chunk_id''], r[''score''])\n print(r[''content''][:200])\n print(''---'')","stdout":"17\nNone 360a55ef-1fae-45c7-9506-d681bf7d2642 - 0.032786883413791656\nPhase 2: Label selection and guideline. We reviewed the collected documents and identified - the most common structural features they exhibit. This was achieved by identifying recurrent layout elements \n---\nNone - f71a4730-2d9c-4317-9e33-2488adda8f3d 0.0317540317773819\n\nTable 1: DocLayNet dataset overview. Along with the frequency - of each class label, we present the relative occurrence (as % of row \"Total\") in the train, test and validation - sets. The inter-annotator\n---\nNone 994f8aeb-bdf3-434d-9b2e-69d4a6a9c623 0.03015873022377491\n$_{Affiliation}$, - as seen in DocBank, are often only distinguishable by discriminating on\nPreparation work included uploading and - parsing the sourced PDF documents in the Corpus Conversion Service (CC\n---\nNone 70a1c951-bc93-4302-95ca-bdbb832f3cf9 - 0.029462365433573723\nPhase 1: Data selection and preparation. Our inclusion criteria for documents were described - in Section 3. A large effort went into ensuring that all documents are free to use. The data sources includ\n---\nNone - 916ed8c5-d868-4064-a459-1f2cc704df4e 0.028371628373861313\nmAP @ 0.5-0.95 (%).Sci = 89-94. Total, triple inter-annotator - mAP @ 0.5-0.95 (%).Law = 86-91. Total, triple inter-annotator mAP @ 0.5-0.95 (%).Pat = 71-76. Total, triple inter-annotator - mAP @ 0.5-0.95\n---\nNone 41a5b3f5-ff96-4856-9eb9-4695fe28b39c 0.01587301678955555\nCaption, Count = 22524. Caption, - % of Total.Train = 2.04. Caption, % of Total.Test = 1.77. Caption, % of Total.Val = 2.32. Caption, triple inter-annotator - mAP @ 0.5-0.95 (%).All = 84-89. Caption, trip\n---\nNone c359d67a-0809-45bb-bfb0-139817b967fd 0.015625\nPage-footer, - triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 96-98. Page-header, Count = 58022. Page-header, % of Total.Train - = 5.10. Page-header, % of Total.Test = 6.70. Page-header, % of Total.Val =\n---\nNone 50e95bc9-862e-4bdb-9e8c-4d1e38d0eee6 - 0.015384615398943424\n0.5-0.95 (%).Law = 87-94. Section-header, triple inter-annotator mAP @ 0.5-0.95 (%).Pat = - 69-73. Section-header, triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 78-86. Table, Count = 34733. Table, % o\n---\nNone - d40d3add-cd91-4774-9d3e-77c388e0f9a4 0.01515151560306549\n89-93. Text, triple inter-annotator mAP @ 0.5-0.95 (%).Law - = 87-92. Text, triple inter-annotator mAP @ 0.5-0.95 (%).Pat = 71-79. Text, triple inter-annotator mAP @ 0.5-0.95 - (%).Ten = 87-95. Title, Cou\n---\nNone 683d9ab8-6363-4165-9cc7-b89145db3f33 0.014925372786819935\n185660. List-item, - % of Total.Train = 17.19. List-item, % of Total.Test = 13.34. List-item, % of Total.Val = 15.82. List-item, triple - inter-annotator mAP @ 0.5-0.95 (%).All = 87-88. List-item, triple \n---\nNone f5d1d638-002e-42ab-8c35-6fbf834ab435 - 0.014705882407724857\ninter-annotator mAP @ 0.5-0.95 (%).Pat = 66-71. Picture, triple inter-annotator mAP @ 0.5-0.95 - (%).Ten = 59-76. Section-header, Count = 142884. Section-header, % of Total.Train = 12.60. Section-header\n---\nNone - 171eb4b0-e518-4e65-9ef4-5655789dceae 0.014492753893136978\nn/a. Footnote, Count = 6318. Footnote, % of Total.Train - = 0.60. Footnote, % of Total.Test = 0.31. Footnote, % of Total.Val = 0.58. Footnote, triple inter-annotator mAP - @ 0.5-0.95 (%).All = 83-91. Foot\n---\nNone 64741ed7-d686-4e6c-b8e1-462a51282ecc 0.014084506779909134\n(%).Sci = - 94-100. Title, triple inter-annotator mAP @ 0.5-0.95 (%).Law = 82-96. Title, triple inter-annotator mAP @ 0.5-0.95 - (%).Pat = 68-79. Title, triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 24-\n---\nNone d8771358-ce2e-407f-af30-ee410ed8a579 - 0.013888888992369175\n0.5-0.95 (%).Pat = 97-99. Page-header, triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 81-86. - Picture, Count = 45976. Picture, % of Total.Train = 4.21. Picture, % of Total.Test = 2.78. Picture, % of T\n---\nNone - c2289d48-99bf-4400-ac39-155b764485db 0.013698630034923553\ninter-annotator mAP @ 0.5-0.95 (%).Law = 58-80. Table, - triple inter-annotator mAP @ 0.5-0.95 (%).Pat = 79-84. Table, triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 70-85. - Text, Count = 510377. Text, \n---\nNone 6a21d905-7202-4ed9-9241-32934c09f36e 0.013513513840734959\n0.5-0.95 (%).Ten - = 93-95. Page-footer, Count = 70878. Page-footer, % of Total.Train = 6.51. Page-footer, % of Total.Test = 5.58. - Page-footer, % of Total.Val = 6.00. Page-footer, triple inter-annotator\n---\nNone 610e94fb-1051-4e09-b94b-98cd485e551f - 0.01315789483487606\n= 82-97. Formula, Count = 25027. Formula, % of Total.Train = 2.25. Formula, % of Total.Test - = 1.90. Formula, % of Total.Val = 2.96. Formula, triple inter-annotator mAP @ 0.5-0.95 (%).All = 83-85. Form\n---\n","stderr":"","success":true}' - role: tool - tool_call_id: call_o2vpon36 - - content: |- - - Chapters mention labels like "page-header", "page-footer", "section-header", "picture", "table", "text", "title", "list-item", "footnote", "formula". Also maybe "Caption". We'll extract unique set. - - role: assistant - tool_calls: - - function: - arguments: '{"code":"labels=set()\nfor r in results:\n if ''label'' in r[''content'']:\n pass\n","stdout":"","stderr":"FileNotFoundError: - [Errno 2] No such file or directory: ''results''\n","success":false}' - name: execute_code - id: call_1mr5hrfc - type: function - - content: |- - 3 validation errors: - ```json - [ - { - "type": "extra_forbidden", - "loc": [ - "stdout" - ], - "msg": "Extra inputs are not permitted", - "input": "" - }, - { - "type": "extra_forbidden", - "loc": [ - "stderr" - ], - "msg": "Extra inputs are not permitted", - "input": "FileNotFoundError: [Errno 2] No such file or directory: 'results'\n" - }, - { - "type": "extra_forbidden", - "loc": [ - "success" - ], - "msg": "Extra inputs are not permitted", - "input": false - } - ] - ``` - - Fix the errors and try again. - role: tool - tool_call_id: call_1mr5hrfc - - content: |- - - Need to store results variable outside. Let's re-run search and store. - - role: assistant - tool_calls: - - function: - arguments: '{"code":"results = search(''document element types labels'', limit=20)\nlabels=set()\nimport re\nfor - r in results:\n # extract words like \"Caption\", \"Formula\", etc by regex\n for m in re.finditer(r\"\\b([A-Za-z-]+)\\b\", - r[''content'']):\n labels.add(m.group(1))\nprint(sorted(labels)[:30])\nprint(''count'', len(labels))"}' - name: execute_code - id: call_u9fkvji9 - type: function - - content: '{"code":"results = search(''document element types labels'', limit=20)\nlabels=set()\nimport re\nfor r in - results:\n # extract words like \"Caption\", \"Formula\", etc by regex\n for m in re.finditer(r\"\\b([A-Za-z-]+)\\b\", - r[''content'']):\n labels.add(m.group(1))\nprint(sorted(labels)[:30])\nprint(''count'', len(labels))","stdout":"[''-'', - ''A'', ''Abstract'', ''Affiliation'', ''All'', ''Along'', ''Articles'', ''Author'', ''CCS'', ''Caption'', ''Conversion'', - ''Corpus'', ''Count'', ''Critical'', ''Data'', ''DocBank'', ''DocLayNet'', ''Figure'', ''Fin'', ''Footnote'', ''For'', - ''Formula'', ''Label'', ''Labels'', ''Law'', ''List-item'', ''Man'', ''Our'', ''PDF'', ''Page'']\ncount 295\n","stderr":"","success":true}' - role: tool - tool_call_id: call_u9fkvji9 - model: gpt-oss - reasoning_effort: low - stream: false - tool_choice: auto - tools: - - function: - description: |- - Execute Python code in a Docker-sandboxed environment. - - The code has access to haiku.rag functions (search, list_documents, - get_document, get_docling_document, llm) and any Python standard - library module. - - Use print() to output results. - - Structured result with success status, stdout, and stderr. - - name: execute_code - parameters: - additionalProperties: false - properties: - code: - description: Python code to execute. - type: string - required: - - code - type: object - strict: true - type: function - - function: - description: Result from RLM agent execution. - name: final_result - parameters: - additionalProperties: false - properties: - answer: - description: The answer to the user's question - type: string - program: - description: The final consolidated program - type: string - required: - - answer - - program - type: object - strict: true - type: function - uri: http://localhost:11434/v1/chat/completions + encoding_format: base64 + input: + - label + model: qwen3-embedding:4b + uri: http://localhost:11434/v1/embeddings response: headers: - content-length: - - '732' content-type: - application/json + transfer-encoding: + - chunked parsed_body: - choices: - - finish_reason: tool_calls + data: + - embedding: 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 index: 0 - message: - content: '' - reasoning: Too generic. Instead, use label list from dataset. Let's search for "label\" in dataset info. - role: assistant - tool_calls: - - function: - arguments: '{"code":"results = search(''label\\n'', limit=20)\nprint(len(results))\nfor r in results[:3]:\n print(r[''document_title''], - r[''chunk_id''], r[''score''])\n print(r[''content''][:200])\n print(''---'')"}' - name: execute_code - id: call_8m1hy7q3 - index: 0 - type: function - created: 1770373437 - id: chatcmpl-989 - model: gpt-oss - object: chat.completion - system_fingerprint: fp_ollama + object: embedding + model: qwen3-embedding:4b + object: list usage: - completion_tokens: 106 - prompt_tokens: 4481 - total_tokens: 4587 + prompt_tokens: 2 + total_tokens: 2 status: code: 200 message: OK @@ -1729,7 +1053,7 @@ interactions: connection: - keep-alive content-length: - - '17584' + - '15687' content-type: - application/json host: @@ -1742,32 +1066,33 @@ interactions: IMPORTANT: You MUST use the `execute_code` tool to run Python code. The functions described below are ONLY available inside the execute_code tool - you cannot access them any other way. Always execute code to answer questions; do not just describe what code would do. - CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just use them directly: - - search("query") ✓ CORRECT + CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just call them with `await`: + - results = await search("query") ✓ CORRECT - from haiku.rag import search ✗ WRONG - will fail + - results = search("query") ✗ WRONG - must use await - You have access to a sandboxed Python environment with these haiku.rag functions (use them directly, no imports needed): + You have access to a sandboxed Python interpreter with these haiku.rag functions (use them directly with `await`, no imports needed): ## Available Functions - ### search(query, limit=10) -> list[dict] + ### await search(query, limit=10) -> list[dict] Search the knowledge base using hybrid search (vector + full-text). Returns list of dicts with keys: chunk_id, content, document_id, document_title, document_uri, score, page_numbers, headings - ### list_documents(limit=10, offset=0) -> list[dict] + ### await list_documents(limit=10, offset=0) -> list[dict] List available documents in the knowledge base. Returns list of dicts with keys: id, title, uri, created_at - ### get_document(id_or_title) -> str | None + ### await get_document(id_or_title) -> str | None Get the full text content of a document by ID, title, or URI. Returns the document content as a string, or None if not found. - ### get_docling_document(id_or_title) -> DoclingDocument | None - Get the structured DoclingDocument object for advanced analysis. - Returns a DoclingDocument object, or None if not found. - See "DoclingDocument API" section below for how to use it. + ### await get_chunk(chunk_id) -> dict | None + Get a specific chunk by its ID (from search results). + Returns dict with keys: chunk_id, content, document_id, document_title, headings, page_numbers, labels + Use this to retrieve full chunk details and metadata for citation. - ### llm(prompt) -> str + ### await llm(prompt) -> str Call an LLM directly with the given prompt. Returns the response as a string. Use this for classification, summarization, extraction, or any task where you already have the content and just need LLM reasoning. @@ -1782,93 +1107,64 @@ interactions: ``` Check if it exists with: `if 'documents' in dir(): ...` - ## Standard Library Modules - You can import any Python standard library module. + ## Available Python Features + + The interpreter supports: variables, arithmetic, strings, f-strings, lists, dicts, tuples, sets, loops, conditionals, comprehensions, functions, async/await, `map()`, `sorted()`/`.sort(key=...)`, try/except, and the `json` module. + + Not supported: imports (other than `json`), class definitions, generators/yield, match statements, decorators, `with` statements. + + For pattern matching or text extraction, use string methods (`str.split`, `str.find`, `str.startswith`, `in` operator) or the `llm()` function. ## Strategy Guide 1. **Explore First**: Start by listing documents or searching to understand what's available. Document names may differ from filenames (e.g., "tbmed593.pdf" might be stored as "TB MED 593" or similar). - 2. **If get_document returns None**: Use `list_documents()` to see actual document titles, or `search()` to find relevant content. + 2. **If get_document returns None**: Use `await list_documents()` to see actual document titles, or `await search()` to find relevant content. 3. **Iterative Refinement**: Run code, examine results, adjust your approach based on what you find. - 4. **Use print() Liberally**: The REPL captures stdout - print intermediate results to see what you're working with. - 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and use collections. - 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use llm(). - 7. **Cite Your Sources**: Track which documents/chunks informed your answer for citation. - - ## DoclingDocument API - - When you call `get_docling_document(id_or_title)`, you get a DoclingDocument object for structured document analysis. - - ### Properties - - `doc.texts` - List of all text items (paragraphs, headings, etc.) - - `doc.tables` - List of all tables - - `doc.pictures` - List of all pictures/figures - - `doc.name` - Document name - - ### Methods - - `doc.iterate_items(with_groups=False)` - Iterate all items with hierarchy level - Returns tuples of (item, level) where level is nesting depth - - `doc.export_to_markdown()` - Export entire document as markdown string - - ### Text Item Properties - - `item.text` - The text content - - `item.label` - Type: title, paragraph, section_header, list_item, etc. (lowercase enum values) - - `item.prov` - Provenance (page numbers, bounding boxes) - - ### Table Access - - `table.data.num_rows`, `table.data.num_cols` - Dimensions - - `table.data.table_cells` - List of TableCell objects - - `cell.text`, `cell.start_row_offset_idx`, `cell.start_col_offset_idx` - - ### Example Usage - ```python - doc = get_docling_document("My Document") - - # Get all headings - headings = [t.text for t in doc.texts if "header" in str(t.label)] - - # Iterate with structure - for item, level in doc.iterate_items(): - print(" " * level + item.text[:50]) - - # Extract table data - for table in doc.tables: - for cell in table.data.table_cells: - print(f"Row {cell.start_row_offset_idx}, Col {cell.start_col_offset_idx}: {cell.text}") - ``` + 4. **Use print() Liberally**: The sandbox captures stdout - print intermediate results to see what you're working with. + 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and data structures. + 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use `await llm()`. + 7. **Cite Your Sources**: Use get_chunk() to retrieve chunk metadata for citations. Track which documents/chunks informed your answer. ## Example Patterns ### Counting documents matching a condition ```python - docs = list_documents(limit=100) + docs = await list_documents(limit=100) count = 0 for doc in docs: - content = get_document(doc['id']) + content = await get_document(doc['id']) if content and 'keyword' in content.lower(): count += 1 print(f"Found in: {doc['title']}") print(f"Total: {count}") ``` - ### Aggregating data across documents + ### Extracting data with llm() ```python - import re numbers = [] - results = search("financial data", limit=20) + results = await search("financial data", limit=20) for r in results: - matches = re.findall(r'\$([\d,]+)', r['content']) - for m in matches: - numbers.append(int(m.replace(',', ''))) - print(f"Average: ${sum(numbers)/len(numbers):,.2f}") + extracted = await llm(f"Extract all dollar amounts from this text as a comma-separated list of numbers (no $ signs): {r['content']}") + for part in extracted.split(','): + part = part.strip().replace(',', '') + if part.isdigit(): + numbers.append(int(part)) + if numbers: + print(f"Average: {sum(numbers) / len(numbers)}") + ``` + + ### Using search results with get_chunk for citations + ```python + results = await search("safety requirements", limit=5) + for r in results: + chunk = await get_chunk(r['chunk_id']) + print(f"From '{chunk['document_title']}', page {chunk['page_numbers']}: {chunk['content'][:100]}") ``` ### Using llm() for classification ```python - # Get document content - content = get_document("Q1 Report") - # Use llm() to classify sentiment - sentiment = llm(f"Classify the sentiment as positive, negative, or mixed: {content}") + content = await get_document("Q1 Report") + sentiment = await llm(f"Classify the sentiment as positive, negative, or mixed: {content}") print(sentiment) ``` @@ -1895,589 +1191,99 @@ interactions: - content: Search for content about document element types or labels. What are all the different document element types mentioned? List them all. role: user - - content: |- - - We need search for "document element types" or "labels". We'll run search. - + - content: null + reasoning: Need to search for document element types or labels. Use search. role: assistant tool_calls: - function: - arguments: '{"code":"results = search(''document element types labels'', limit=20)\nprint(len(results))\nfor r - in results:\n print(r[''document_title''], r[''chunk_id''], r[''score''])\n print(r[''content''][:200])\n print(''---'')"}' + arguments: '{"code":"results = await search(\"document element types\" , limit=20)\nprint(len(results))\nfor r + in results[:5]:\n print(r[''chunk_id''], r[''document_title''], r[''content''][:200])\n"}' name: execute_code - id: call_o2vpon36 + id: call_j7uw3i6n type: function - - content: '{"code":"results = search(''document element types labels'', limit=20)\nprint(len(results))\nfor r in results:\n print(r[''document_title''], - r[''chunk_id''], r[''score''])\n print(r[''content''][:200])\n print(''---'')","stdout":"17\nNone 360a55ef-1fae-45c7-9506-d681bf7d2642 - 0.032786883413791656\nPhase 2: Label selection and guideline. We reviewed the collected documents and identified - the most common structural features they exhibit. This was achieved by identifying recurrent layout elements \n---\nNone - f71a4730-2d9c-4317-9e33-2488adda8f3d 0.0317540317773819\n\nTable 1: DocLayNet dataset overview. Along with the frequency - of each class label, we present the relative occurrence (as % of row \"Total\") in the train, test and validation - sets. The inter-annotator\n---\nNone 994f8aeb-bdf3-434d-9b2e-69d4a6a9c623 0.03015873022377491\n$_{Affiliation}$, - as seen in DocBank, are often only distinguishable by discriminating on\nPreparation work included uploading and - parsing the sourced PDF documents in the Corpus Conversion Service (CC\n---\nNone 70a1c951-bc93-4302-95ca-bdbb832f3cf9 - 0.029462365433573723\nPhase 1: Data selection and preparation. Our inclusion criteria for documents were described - in Section 3. A large effort went into ensuring that all documents are free to use. The data sources includ\n---\nNone - 916ed8c5-d868-4064-a459-1f2cc704df4e 0.028371628373861313\nmAP @ 0.5-0.95 (%).Sci = 89-94. Total, triple inter-annotator - mAP @ 0.5-0.95 (%).Law = 86-91. Total, triple inter-annotator mAP @ 0.5-0.95 (%).Pat = 71-76. Total, triple inter-annotator - mAP @ 0.5-0.95\n---\nNone 41a5b3f5-ff96-4856-9eb9-4695fe28b39c 0.01587301678955555\nCaption, Count = 22524. Caption, - % of Total.Train = 2.04. Caption, % of Total.Test = 1.77. Caption, % of Total.Val = 2.32. Caption, triple inter-annotator - mAP @ 0.5-0.95 (%).All = 84-89. Caption, trip\n---\nNone c359d67a-0809-45bb-bfb0-139817b967fd 0.015625\nPage-footer, - triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 96-98. Page-header, Count = 58022. Page-header, % of Total.Train - = 5.10. Page-header, % of Total.Test = 6.70. Page-header, % of Total.Val =\n---\nNone 50e95bc9-862e-4bdb-9e8c-4d1e38d0eee6 - 0.015384615398943424\n0.5-0.95 (%).Law = 87-94. Section-header, triple inter-annotator mAP @ 0.5-0.95 (%).Pat = - 69-73. Section-header, triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 78-86. Table, Count = 34733. Table, % o\n---\nNone - d40d3add-cd91-4774-9d3e-77c388e0f9a4 0.01515151560306549\n89-93. Text, triple inter-annotator mAP @ 0.5-0.95 (%).Law - = 87-92. Text, triple inter-annotator mAP @ 0.5-0.95 (%).Pat = 71-79. Text, triple inter-annotator mAP @ 0.5-0.95 - (%).Ten = 87-95. Title, Cou\n---\nNone 683d9ab8-6363-4165-9cc7-b89145db3f33 0.014925372786819935\n185660. List-item, - % of Total.Train = 17.19. List-item, % of Total.Test = 13.34. List-item, % of Total.Val = 15.82. List-item, triple - inter-annotator mAP @ 0.5-0.95 (%).All = 87-88. List-item, triple \n---\nNone f5d1d638-002e-42ab-8c35-6fbf834ab435 - 0.014705882407724857\ninter-annotator mAP @ 0.5-0.95 (%).Pat = 66-71. Picture, triple inter-annotator mAP @ 0.5-0.95 - (%).Ten = 59-76. Section-header, Count = 142884. Section-header, % of Total.Train = 12.60. Section-header\n---\nNone - 171eb4b0-e518-4e65-9ef4-5655789dceae 0.014492753893136978\nn/a. Footnote, Count = 6318. Footnote, % of Total.Train - = 0.60. Footnote, % of Total.Test = 0.31. Footnote, % of Total.Val = 0.58. Footnote, triple inter-annotator mAP - @ 0.5-0.95 (%).All = 83-91. Foot\n---\nNone 64741ed7-d686-4e6c-b8e1-462a51282ecc 0.014084506779909134\n(%).Sci = - 94-100. Title, triple inter-annotator mAP @ 0.5-0.95 (%).Law = 82-96. Title, triple inter-annotator mAP @ 0.5-0.95 - (%).Pat = 68-79. Title, triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 24-\n---\nNone d8771358-ce2e-407f-af30-ee410ed8a579 - 0.013888888992369175\n0.5-0.95 (%).Pat = 97-99. Page-header, triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 81-86. - Picture, Count = 45976. Picture, % of Total.Train = 4.21. Picture, % of Total.Test = 2.78. Picture, % of T\n---\nNone - c2289d48-99bf-4400-ac39-155b764485db 0.013698630034923553\ninter-annotator mAP @ 0.5-0.95 (%).Law = 58-80. Table, - triple inter-annotator mAP @ 0.5-0.95 (%).Pat = 79-84. Table, triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 70-85. - Text, Count = 510377. Text, \n---\nNone 6a21d905-7202-4ed9-9241-32934c09f36e 0.013513513840734959\n0.5-0.95 (%).Ten - = 93-95. Page-footer, Count = 70878. Page-footer, % of Total.Train = 6.51. Page-footer, % of Total.Test = 5.58. - Page-footer, % of Total.Val = 6.00. Page-footer, triple inter-annotator\n---\nNone 610e94fb-1051-4e09-b94b-98cd485e551f - 0.01315789483487606\n= 82-97. Formula, Count = 25027. Formula, % of Total.Train = 2.25. Formula, % of Total.Test - = 1.90. Formula, % of Total.Val = 2.96. Formula, triple inter-annotator mAP @ 0.5-0.95 (%).All = 83-85. Form\n---\n","stderr":"","success":true}' + - content: '{"code":"results = await search(\"document element types\" , limit=20)\nprint(len(results))\nfor r in results[:5]:\n print(r[''chunk_id''], + r[''document_title''], r[''content''][:200])\n","stdout":"17\n7c3a912c-a769-49e9-a661-402d7d64ec9f None Phase 2: + Label selection and guideline. We reviewed the collected documents and identified the most common structural features + they exhibit. This was achieved by identifying recurrent layout elements \n62e98936-a694-4cb0-a1a4-409b0121e08d + None Phase 1: Data selection and preparation. Our inclusion criteria for documents were described in Section 3. + A large effort went into ensuring that all documents are free to use. The data sources includ\nd583f0ed-c230-4f8e-988a-86875c0c633d + None $_{Affiliation}$, as seen in DocBank, are often only distinguishable by discriminating on\nPreparation work + included uploading and parsing the sourced PDF documents in the Corpus Conversion Service (CC\n180df00a-2a5b-4e2a-bfcc-4cd67fbc9db0 + None \nTable 1: DocLayNet dataset overview. Along with the frequency of each class label, we present the relative + occurrence (as % of row \"Total\") in the train, test and validation sets. The inter-annotator\nf3c18411-85df-4c72-a784-f39f602745ec + None Page-footer, triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 96-98. Page-header, Count = 58022. Page-header, + % of Total.Train = 5.10. Page-header, % of Total.Test = 6.70. Page-header, % of Total.Val =\n","stderr":"","success":true}' role: tool - tool_call_id: call_o2vpon36 - - content: |- - - Chapters mention labels like "page-header", "page-footer", "section-header", "picture", "table", "text", "title", "list-item", "footnote", "formula". Also maybe "Caption". We'll extract unique set. - + tool_call_id: call_j7uw3i6n + - content: null + reasoning: We need list all document element types or labels mentioned. Likely these are labels in documents like + "Page-header", "Page-footer", etc. Let's search more. role: assistant tool_calls: - function: - arguments: '{"code":"labels=set()\nfor r in results:\n if ''label'' in r[''content'']:\n pass\n","stdout":"","stderr":"FileNotFoundError: - [Errno 2] No such file or directory: ''results''\n","success":false}' + arguments: '{"code":"results = await search(\"label\" , limit=30)\nlabels=set()\nfor r in results:\n # look + for uppercase words separated by hyphens\n import re\n for m in re.findall(r\"([A-Z][A-Za-z]+(?:-[A-Z][A-Za-z]+)*)\", + r[''content'']):\n labels.add(m)\nprint(sorted(labels))\n"}' name: execute_code - id: call_1mr5hrfc + id: call_gzg850pl type: function - - content: |- - 3 validation errors: - ```json - [ - { - "type": "extra_forbidden", - "loc": [ - "stdout" - ], - "msg": "Extra inputs are not permitted", - "input": "" - }, - { - "type": "extra_forbidden", - "loc": [ - "stderr" - ], - "msg": "Extra inputs are not permitted", - "input": "FileNotFoundError: [Errno 2] No such file or directory: 'results'\n" - }, - { - "type": "extra_forbidden", - "loc": [ - "success" - ], - "msg": "Extra inputs are not permitted", - "input": false - } - ] - ``` - - Fix the errors and try again. + - content: '{"code":"results = await search(\"label\" , limit=30)\nlabels=set()\nfor r in results:\n # look for uppercase + words separated by hyphens\n import re\n for m in re.findall(r\"([A-Z][A-Za-z]+(?:-[A-Z][A-Za-z]+)*)\", r[''content'']):\n labels.add(m)\nprint(sorted(labels))\n","stdout":"","stderr":"ModuleNotFoundError: + No module named ''re''","success":false}' role: tool - tool_call_id: call_1mr5hrfc - - content: |- - - Need to store results variable outside. Let's re-run search and store. - + tool_call_id: call_gzg850pl + - content: null + reasoning: Import not allowed. Use string methods. Probably easier to manually inspect. Let's list results. role: assistant tool_calls: - function: - arguments: '{"code":"results = search(''document element types labels'', limit=20)\nlabels=set()\nimport re\nfor - r in results:\n # extract words like \"Caption\", \"Formula\", etc by regex\n for m in re.finditer(r\"\\b([A-Za-z-]+)\\b\", - r[''content'']):\n labels.add(m.group(1))\nprint(sorted(labels)[:30])\nprint(''count'', len(labels))"}' + arguments: '{"code":"results = await search(\"label\" , limit=10)\nfor r in results:\n print(''---'')\n print(r[''chunk_id''], + r[''document_title''])\n print(r[''content''][:400])\n"}' name: execute_code - id: call_u9fkvji9 + id: call_4s9httx6 type: function - - content: '{"code":"results = search(''document element types labels'', limit=20)\nlabels=set()\nimport re\nfor r in - results:\n # extract words like \"Caption\", \"Formula\", etc by regex\n for m in re.finditer(r\"\\b([A-Za-z-]+)\\b\", - r[''content'']):\n labels.add(m.group(1))\nprint(sorted(labels)[:30])\nprint(''count'', len(labels))","stdout":"[''-'', - ''A'', ''Abstract'', ''Affiliation'', ''All'', ''Along'', ''Articles'', ''Author'', ''CCS'', ''Caption'', ''Conversion'', - ''Corpus'', ''Count'', ''Critical'', ''Data'', ''DocBank'', ''DocLayNet'', ''Figure'', ''Fin'', ''Footnote'', ''For'', - ''Formula'', ''Label'', ''Labels'', ''Law'', ''List-item'', ''Man'', ''Our'', ''PDF'', ''Page'']\ncount 295\n","stderr":"","success":true}' - role: tool - tool_call_id: call_u9fkvji9 - - content: |- - - Too generic. Instead, use label list from dataset. Let's search for "label\" in dataset info. - - role: assistant - tool_calls: - - function: - arguments: '{"code":"results = search(''label\\n'', limit=20)\nprint(len(results))\nfor r in results[:3]:\n print(r[''document_title''], - r[''chunk_id''], r[''score''])\n print(r[''content''][:200])\n print(''---'')"}' - name: execute_code - id: call_8m1hy7q3 - type: function - - content: '{"code":"results = search(''label\\n'', limit=20)\nprint(len(results))\nfor r in results[:3]:\n print(r[''document_title''], - r[''chunk_id''], r[''score''])\n print(r[''content''][:200])\n print(''---'')","stdout":"17\nNone 360a55ef-1fae-45c7-9506-d681bf7d2642 - 0.032786883413791656\nPhase 2: Label selection and guideline. We reviewed the collected documents and identified - the most common structural features they exhibit. This was achieved by identifying recurrent layout elements \n---\nNone - f71a4730-2d9c-4317-9e33-2488adda8f3d 0.03151364624500275\n\nTable 1: DocLayNet dataset overview. Along with the - frequency of each class label, we present the relative occurrence (as % of row \"Total\") in the train, test and - validation sets. The inter-annotator\n---\nNone 916ed8c5-d868-4064-a459-1f2cc704df4e 0.02903090976178646\nmAP @ - 0.5-0.95 (%).Sci = 89-94. Total, triple inter-annotator mAP @ 0.5-0.95 (%).Law = 86-91. Total, triple inter-annotator - mAP @ 0.5-0.95 (%).Pat = 71-76. Total, triple inter-annotator mAP @ 0.5-0.95\n---\n","stderr":"","success":true}' - role: tool - tool_call_id: call_8m1hy7q3 - model: gpt-oss - reasoning_effort: low - stream: false - tool_choice: auto - tools: - - function: - description: |- - Execute Python code in a Docker-sandboxed environment. - - The code has access to haiku.rag functions (search, list_documents, - get_document, get_docling_document, llm) and any Python standard - library module. - - Use print() to output results. - - Structured result with success status, stdout, and stderr. - - name: execute_code - parameters: - additionalProperties: false - properties: - code: - description: Python code to execute. - type: string - required: - - code - type: object - strict: true - type: function - - function: - description: Result from RLM agent execution. - name: final_result - parameters: - additionalProperties: false - properties: - answer: - description: The answer to the user's question - type: string - program: - description: The final consolidated program - type: string - required: - - answer - - program - type: object - strict: true - type: function - uri: http://localhost:11434/v1/chat/completions - response: - headers: - content-length: - - '790' - content-type: - - application/json - parsed_body: - choices: - - finish_reason: tool_calls - index: 0 - message: - content: '' - reasoning: 'Better: list of element types from dataset: title, heading, paragraph, list_item, table, picture, formula, - footnote, caption, page_header, page_footer, section_header. Let''s extract from dataset table. Search for "Table - 1:" maybe content lists types.' - role: assistant - tool_calls: - - function: - arguments: '{"code":"results = search(''Table 1: DocLayNet'', limit=5)\nfor r in results:\n print(r[''content''][:500])"}' - name: execute_code - id: call_jyyqj1qx - index: 0 - type: function - created: 1770373441 - id: chatcmpl-319 - model: gpt-oss - object: chat.completion - system_fingerprint: fp_ollama - usage: - completion_tokens: 111 - prompt_tokens: 4968 - total_tokens: 5079 - status: - code: 200 - message: OK -- request: - headers: - accept: - - application/json - accept-encoding: - - gzip, deflate, zstd - connection: - - keep-alive - content-length: - - '20704' - content-type: - - application/json - host: - - localhost:11434 - method: POST - parsed_body: - messages: - - content: |- - You are a Recursive Language Model (RLM) agent that solves complex research questions by writing and executing Python code. - - IMPORTANT: You MUST use the `execute_code` tool to run Python code. The functions described below are ONLY available inside the execute_code tool - you cannot access them any other way. Always execute code to answer questions; do not just describe what code would do. - - CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just use them directly: - - search("query") ✓ CORRECT - - from haiku.rag import search ✗ WRONG - will fail - - You have access to a sandboxed Python environment with these haiku.rag functions (use them directly, no imports needed): - - ## Available Functions - - ### search(query, limit=10) -> list[dict] - Search the knowledge base using hybrid search (vector + full-text). - Returns list of dicts with keys: chunk_id, content, document_id, document_title, document_uri, score, page_numbers, headings - - ### list_documents(limit=10, offset=0) -> list[dict] - List available documents in the knowledge base. - Returns list of dicts with keys: id, title, uri, created_at - - ### get_document(id_or_title) -> str | None - Get the full text content of a document by ID, title, or URI. - Returns the document content as a string, or None if not found. - - ### get_docling_document(id_or_title) -> DoclingDocument | None - Get the structured DoclingDocument object for advanced analysis. - Returns a DoclingDocument object, or None if not found. - See "DoclingDocument API" section below for how to use it. - - ### llm(prompt) -> str - Call an LLM directly with the given prompt. Returns the response as a string. - Use this for classification, summarization, extraction, or any task where you - already have the content and just need LLM reasoning. - - ## Pre-loaded Documents Variable - - If documents were pre-loaded for this session, a `documents` variable is available: - ```python - # documents is a list of dicts with keys: id, title, uri, content - for doc in documents: - print(doc['title'], len(doc['content'])) - ``` - Check if it exists with: `if 'documents' in dir(): ...` - - ## Standard Library Modules - You can import any Python standard library module. - - ## Strategy Guide - - 1. **Explore First**: Start by listing documents or searching to understand what's available. Document names may differ from filenames (e.g., "tbmed593.pdf" might be stored as "TB MED 593" or similar). - 2. **If get_document returns None**: Use `list_documents()` to see actual document titles, or `search()` to find relevant content. - 3. **Iterative Refinement**: Run code, examine results, adjust your approach based on what you find. - 4. **Use print() Liberally**: The REPL captures stdout - print intermediate results to see what you're working with. - 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and use collections. - 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use llm(). - 7. **Cite Your Sources**: Track which documents/chunks informed your answer for citation. - - ## DoclingDocument API - - When you call `get_docling_document(id_or_title)`, you get a DoclingDocument object for structured document analysis. - - ### Properties - - `doc.texts` - List of all text items (paragraphs, headings, etc.) - - `doc.tables` - List of all tables - - `doc.pictures` - List of all pictures/figures - - `doc.name` - Document name - - ### Methods - - `doc.iterate_items(with_groups=False)` - Iterate all items with hierarchy level - Returns tuples of (item, level) where level is nesting depth - - `doc.export_to_markdown()` - Export entire document as markdown string - - ### Text Item Properties - - `item.text` - The text content - - `item.label` - Type: title, paragraph, section_header, list_item, etc. (lowercase enum values) - - `item.prov` - Provenance (page numbers, bounding boxes) - - ### Table Access - - `table.data.num_rows`, `table.data.num_cols` - Dimensions - - `table.data.table_cells` - List of TableCell objects - - `cell.text`, `cell.start_row_offset_idx`, `cell.start_col_offset_idx` - - ### Example Usage - ```python - doc = get_docling_document("My Document") - - # Get all headings - headings = [t.text for t in doc.texts if "header" in str(t.label)] - - # Iterate with structure - for item, level in doc.iterate_items(): - print(" " * level + item.text[:50]) - - # Extract table data - for table in doc.tables: - for cell in table.data.table_cells: - print(f"Row {cell.start_row_offset_idx}, Col {cell.start_col_offset_idx}: {cell.text}") - ``` - - ## Example Patterns - - ### Counting documents matching a condition - ```python - docs = list_documents(limit=100) - count = 0 - for doc in docs: - content = get_document(doc['id']) - if content and 'keyword' in content.lower(): - count += 1 - print(f"Found in: {doc['title']}") - print(f"Total: {count}") - ``` - - ### Aggregating data across documents - ```python - import re - numbers = [] - results = search("financial data", limit=20) - for r in results: - matches = re.findall(r'\$([\d,]+)', r['content']) - for m in matches: - numbers.append(int(m.replace(',', ''))) - print(f"Average: ${sum(numbers)/len(numbers):,.2f}") - ``` - - ### Using llm() for classification - ```python - # Get document content - content = get_document("Q1 Report") - # Use llm() to classify sentiment - sentiment = llm(f"Classify the sentiment as positive, negative, or mixed: {content}") - print(sentiment) - ``` - - ## Workflow - - 1. **ALWAYS start by using execute_code** to explore the knowledge base - 2. Run multiple code blocks as needed to gather information - 3. After collecting data, provide your final answer - - ## Output Format - - CRITICAL: Your final response MUST be valid JSON matching this exact schema: - ```json - {"answer": "Your complete answer here as a string", "program": "Your final consolidated program here as a string"} - ``` - - - `answer`: A clear answer to the user's question with key findings and references to specific documents/chunks. - - `program`: A single, self-contained Python program that produces the answer. Consolidate your exploratory code executions into one clean script. - - Do NOT return arbitrary JSON structures. Always use the exact format: {"answer": "...", "program": "..."} - - CRITICAL: You MUST call execute_code at least once before providing your answer. Never give up without trying to execute code first. - role: system - - content: Search for content about document element types or labels. What are all the different document element types - mentioned? List them all. - role: user - - content: |- - - We need search for "document element types" or "labels". We'll run search. - - role: assistant - tool_calls: - - function: - arguments: '{"code":"results = search(''document element types labels'', limit=20)\nprint(len(results))\nfor r - in results:\n print(r[''document_title''], r[''chunk_id''], r[''score''])\n print(r[''content''][:200])\n print(''---'')"}' - name: execute_code - id: call_o2vpon36 - type: function - - content: '{"code":"results = search(''document element types labels'', limit=20)\nprint(len(results))\nfor r in results:\n print(r[''document_title''], - r[''chunk_id''], r[''score''])\n print(r[''content''][:200])\n print(''---'')","stdout":"17\nNone 360a55ef-1fae-45c7-9506-d681bf7d2642 - 0.032786883413791656\nPhase 2: Label selection and guideline. We reviewed the collected documents and identified - the most common structural features they exhibit. This was achieved by identifying recurrent layout elements \n---\nNone - f71a4730-2d9c-4317-9e33-2488adda8f3d 0.0317540317773819\n\nTable 1: DocLayNet dataset overview. Along with the frequency - of each class label, we present the relative occurrence (as % of row \"Total\") in the train, test and validation - sets. The inter-annotator\n---\nNone 994f8aeb-bdf3-434d-9b2e-69d4a6a9c623 0.03015873022377491\n$_{Affiliation}$, - as seen in DocBank, are often only distinguishable by discriminating on\nPreparation work included uploading and - parsing the sourced PDF documents in the Corpus Conversion Service (CC\n---\nNone 70a1c951-bc93-4302-95ca-bdbb832f3cf9 - 0.029462365433573723\nPhase 1: Data selection and preparation. Our inclusion criteria for documents were described - in Section 3. A large effort went into ensuring that all documents are free to use. The data sources includ\n---\nNone - 916ed8c5-d868-4064-a459-1f2cc704df4e 0.028371628373861313\nmAP @ 0.5-0.95 (%).Sci = 89-94. Total, triple inter-annotator - mAP @ 0.5-0.95 (%).Law = 86-91. Total, triple inter-annotator mAP @ 0.5-0.95 (%).Pat = 71-76. Total, triple inter-annotator - mAP @ 0.5-0.95\n---\nNone 41a5b3f5-ff96-4856-9eb9-4695fe28b39c 0.01587301678955555\nCaption, Count = 22524. Caption, - % of Total.Train = 2.04. Caption, % of Total.Test = 1.77. Caption, % of Total.Val = 2.32. Caption, triple inter-annotator - mAP @ 0.5-0.95 (%).All = 84-89. Caption, trip\n---\nNone c359d67a-0809-45bb-bfb0-139817b967fd 0.015625\nPage-footer, - triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 96-98. Page-header, Count = 58022. Page-header, % of Total.Train - = 5.10. Page-header, % of Total.Test = 6.70. Page-header, % of Total.Val =\n---\nNone 50e95bc9-862e-4bdb-9e8c-4d1e38d0eee6 - 0.015384615398943424\n0.5-0.95 (%).Law = 87-94. Section-header, triple inter-annotator mAP @ 0.5-0.95 (%).Pat = - 69-73. Section-header, triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 78-86. Table, Count = 34733. Table, % o\n---\nNone - d40d3add-cd91-4774-9d3e-77c388e0f9a4 0.01515151560306549\n89-93. Text, triple inter-annotator mAP @ 0.5-0.95 (%).Law - = 87-92. Text, triple inter-annotator mAP @ 0.5-0.95 (%).Pat = 71-79. Text, triple inter-annotator mAP @ 0.5-0.95 - (%).Ten = 87-95. Title, Cou\n---\nNone 683d9ab8-6363-4165-9cc7-b89145db3f33 0.014925372786819935\n185660. List-item, - % of Total.Train = 17.19. List-item, % of Total.Test = 13.34. List-item, % of Total.Val = 15.82. List-item, triple - inter-annotator mAP @ 0.5-0.95 (%).All = 87-88. List-item, triple \n---\nNone f5d1d638-002e-42ab-8c35-6fbf834ab435 - 0.014705882407724857\ninter-annotator mAP @ 0.5-0.95 (%).Pat = 66-71. Picture, triple inter-annotator mAP @ 0.5-0.95 - (%).Ten = 59-76. Section-header, Count = 142884. Section-header, % of Total.Train = 12.60. Section-header\n---\nNone - 171eb4b0-e518-4e65-9ef4-5655789dceae 0.014492753893136978\nn/a. Footnote, Count = 6318. Footnote, % of Total.Train - = 0.60. Footnote, % of Total.Test = 0.31. Footnote, % of Total.Val = 0.58. Footnote, triple inter-annotator mAP - @ 0.5-0.95 (%).All = 83-91. Foot\n---\nNone 64741ed7-d686-4e6c-b8e1-462a51282ecc 0.014084506779909134\n(%).Sci = - 94-100. Title, triple inter-annotator mAP @ 0.5-0.95 (%).Law = 82-96. Title, triple inter-annotator mAP @ 0.5-0.95 - (%).Pat = 68-79. Title, triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 24-\n---\nNone d8771358-ce2e-407f-af30-ee410ed8a579 - 0.013888888992369175\n0.5-0.95 (%).Pat = 97-99. Page-header, triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 81-86. - Picture, Count = 45976. Picture, % of Total.Train = 4.21. Picture, % of Total.Test = 2.78. Picture, % of T\n---\nNone - c2289d48-99bf-4400-ac39-155b764485db 0.013698630034923553\ninter-annotator mAP @ 0.5-0.95 (%).Law = 58-80. Table, - triple inter-annotator mAP @ 0.5-0.95 (%).Pat = 79-84. Table, triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 70-85. - Text, Count = 510377. Text, \n---\nNone 6a21d905-7202-4ed9-9241-32934c09f36e 0.013513513840734959\n0.5-0.95 (%).Ten - = 93-95. Page-footer, Count = 70878. Page-footer, % of Total.Train = 6.51. Page-footer, % of Total.Test = 5.58. - Page-footer, % of Total.Val = 6.00. Page-footer, triple inter-annotator\n---\nNone 610e94fb-1051-4e09-b94b-98cd485e551f - 0.01315789483487606\n= 82-97. Formula, Count = 25027. Formula, % of Total.Train = 2.25. Formula, % of Total.Test - = 1.90. Formula, % of Total.Val = 2.96. Formula, triple inter-annotator mAP @ 0.5-0.95 (%).All = 83-85. Form\n---\n","stderr":"","success":true}' - role: tool - tool_call_id: call_o2vpon36 - - content: |- - - Chapters mention labels like "page-header", "page-footer", "section-header", "picture", "table", "text", "title", "list-item", "footnote", "formula". Also maybe "Caption". We'll extract unique set. - - role: assistant - tool_calls: - - function: - arguments: '{"code":"labels=set()\nfor r in results:\n if ''label'' in r[''content'']:\n pass\n","stdout":"","stderr":"FileNotFoundError: - [Errno 2] No such file or directory: ''results''\n","success":false}' - name: execute_code - id: call_1mr5hrfc - type: function - - content: |- - 3 validation errors: - ```json - [ - { - "type": "extra_forbidden", - "loc": [ - "stdout" - ], - "msg": "Extra inputs are not permitted", - "input": "" - }, - { - "type": "extra_forbidden", - "loc": [ - "stderr" - ], - "msg": "Extra inputs are not permitted", - "input": "FileNotFoundError: [Errno 2] No such file or directory: 'results'\n" - }, - { - "type": "extra_forbidden", - "loc": [ - "success" - ], - "msg": "Extra inputs are not permitted", - "input": false - } - ] - ``` - - Fix the errors and try again. - role: tool - tool_call_id: call_1mr5hrfc - - content: |- - - Need to store results variable outside. Let's re-run search and store. - - role: assistant - tool_calls: - - function: - arguments: '{"code":"results = search(''document element types labels'', limit=20)\nlabels=set()\nimport re\nfor - r in results:\n # extract words like \"Caption\", \"Formula\", etc by regex\n for m in re.finditer(r\"\\b([A-Za-z-]+)\\b\", - r[''content'']):\n labels.add(m.group(1))\nprint(sorted(labels)[:30])\nprint(''count'', len(labels))"}' - name: execute_code - id: call_u9fkvji9 - type: function - - content: '{"code":"results = search(''document element types labels'', limit=20)\nlabels=set()\nimport re\nfor r in - results:\n # extract words like \"Caption\", \"Formula\", etc by regex\n for m in re.finditer(r\"\\b([A-Za-z-]+)\\b\", - r[''content'']):\n labels.add(m.group(1))\nprint(sorted(labels)[:30])\nprint(''count'', len(labels))","stdout":"[''-'', - ''A'', ''Abstract'', ''Affiliation'', ''All'', ''Along'', ''Articles'', ''Author'', ''CCS'', ''Caption'', ''Conversion'', - ''Corpus'', ''Count'', ''Critical'', ''Data'', ''DocBank'', ''DocLayNet'', ''Figure'', ''Fin'', ''Footnote'', ''For'', - ''Formula'', ''Label'', ''Labels'', ''Law'', ''List-item'', ''Man'', ''Our'', ''PDF'', ''Page'']\ncount 295\n","stderr":"","success":true}' - role: tool - tool_call_id: call_u9fkvji9 - - content: |- - - Too generic. Instead, use label list from dataset. Let's search for "label\" in dataset info. - - role: assistant - tool_calls: - - function: - arguments: '{"code":"results = search(''label\\n'', limit=20)\nprint(len(results))\nfor r in results[:3]:\n print(r[''document_title''], - r[''chunk_id''], r[''score''])\n print(r[''content''][:200])\n print(''---'')"}' - name: execute_code - id: call_8m1hy7q3 - type: function - - content: '{"code":"results = search(''label\\n'', limit=20)\nprint(len(results))\nfor r in results[:3]:\n print(r[''document_title''], - r[''chunk_id''], r[''score''])\n print(r[''content''][:200])\n print(''---'')","stdout":"17\nNone 360a55ef-1fae-45c7-9506-d681bf7d2642 - 0.032786883413791656\nPhase 2: Label selection and guideline. We reviewed the collected documents and identified - the most common structural features they exhibit. This was achieved by identifying recurrent layout elements \n---\nNone - f71a4730-2d9c-4317-9e33-2488adda8f3d 0.03151364624500275\n\nTable 1: DocLayNet dataset overview. Along with the - frequency of each class label, we present the relative occurrence (as % of row \"Total\") in the train, test and - validation sets. The inter-annotator\n---\nNone 916ed8c5-d868-4064-a459-1f2cc704df4e 0.02903090976178646\nmAP @ - 0.5-0.95 (%).Sci = 89-94. Total, triple inter-annotator mAP @ 0.5-0.95 (%).Law = 86-91. Total, triple inter-annotator - mAP @ 0.5-0.95 (%).Pat = 71-76. Total, triple inter-annotator mAP @ 0.5-0.95\n---\n","stderr":"","success":true}' - role: tool - tool_call_id: call_8m1hy7q3 - - content: |- - - Better: list of element types from dataset: title, heading, paragraph, list_item, table, picture, formula, footnote, caption, page_header, page_footer, section_header. Let's extract from dataset table. Search for "Table 1:" maybe content lists types. - - role: assistant - tool_calls: - - function: - arguments: '{"code":"results = search(''Table 1: DocLayNet'', limit=5)\nfor r in results:\n print(r[''content''][:500])"}' - name: execute_code - id: call_jyyqj1qx - type: function - - content: '{"code":"results = search(''Table 1: DocLayNet'', limit=5)\nfor r in results:\n print(r[''content''][:500])","stdout":"\nTable + - content: '{"code":"results = await search(\"label\" , limit=10)\nfor r in results:\n print(''---'')\n print(r[''chunk_id''], + r[''document_title''])\n print(r[''content''][:400])\n","stdout":"---\n7c3a912c-a769-49e9-a661-402d7d64ec9f None\nPhase + 2: Label selection and guideline. We reviewed the collected documents and identified the most common structural + features they exhibit. This was achieved by identifying recurrent layout elements and lead us to the definition + of 11 distinct class labels. These 11 class labels are $_{Caption}$, $_{Footnote}$, $_{Formula}$, $_{List-item}$, + Page-$_{footer}$, $_{Page-header}$, $_{Picture}$, $_{Sec\n---\n180df00a-2a5b-4e2a-bfcc-4cd67fbc9db0 None\n\nTable 1: DocLayNet dataset overview. Along with the frequency of each class label, we present the relative occurrence (as % of row \"Total\") in the train, test and validation sets. The inter-annotator agreement is computed as the mAP@0.5-0.95 metric between pairwise annotations from the triple-annotated pages, from which we obtain accuracy - ranges.\nPhase 2: Label selection and guideline. We reviewed the collected documents and identified the most common - structural features they exhibit. This was achieved by identifying recurrent layout elements and lead us to the - definition of 11 distinct class labels. These 11 class labels are $_{Caption}$, $_{Footnote}$, $_{Formula}$, $_{List-item}$, - Page-$_{footer}$, $_{Page-header}$, $_{Picture}$, $_{Section-header}$, $_{Table}$, $_{Text}$, and $_{Title}$. Critical - factors that were considered for the \n0.5-0.95 (%).Law = 87-94. Section-header, triple inter-annotator mAP @ 0.5-0.95 - (%).Pat = 69-73. Section-header, triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 78-86. Table, Count = 34733. Table, - % of Total.Train = 3.20. Table, % of Total.Test = 2.27. Table, % of Total.Val = 3.60. Table, triple inter-annotator - mAP @ 0.5-0.95 (%).All = 77-81. Table, triple inter-annotator mAP @ 0.5-0.95 (%).Fin = 75-80. Table, triple inter-annotator - mAP @ 0.5-0.95 (%).Man = 83-86. Table, triple inter-annotator m\ninter-annotator mAP @ 0.5-0.95 (%).Law = 58-80. - Table, triple inter-annotator mAP @ 0.5-0.95 (%).Pat = 79-84. Table, triple inter-annotator mAP @ 0.5-0.95 (%).Ten - = 70-85. Text, Count = 510377. Text, % of Total.Train = 45.82. Text, % of Total.Test = 49.28. Text, % of Total.Val - = 45.00. Text, triple inter-annotator mAP @ 0.5-0.95 (%).All = 84-86. Text, triple inter-annotator mAP @ 0.5-0.95 - (%).Fin = 81-86. Text, triple inter-annotator mAP @ 0.5-0.95 (%).Man = 88-93. Text, triple inter-annotator m\nPage-footer, - triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 96-98. Page-header, Count = 58022. Page-header, % of Total.Train - = 5.10. Page-header, % of Total.Test = 6.70. Page-header, % of Total.Val = 5.06. Page-header, triple inter-annotator - mAP @ 0.5-0.95 (%).All = 85-89. Page-header, triple inter-annotator mAP @ 0.5-0.95 (%).Fin = 66-76. Page-header, - triple inter-annotator mAP @ 0.5-0.95 (%).Man = 90-94. Page-header, triple inter-annotator mAP @ 0.5-0.95 (%).Sci - = 98-100. Page-header, triple \n","stderr":"","success":true}' + ranges.\n---\n5c6655e2-40d8-4ea8-9741-de6a98895449 None\nCaption, Count = 22524. Caption, % of Total.Train = 2.04. + Caption, % of Total.Test = 1.77. Caption, % of Total.Val = 2.32. Caption, triple inter-annotator mAP @ 0.5-0.95 + (%).All = 84-89. Caption, triple inter-annotator mAP @ 0.5-0.95 (%).Fin = 40-61. Caption, triple inter-annotator + mAP @ 0.5-0.95 (%).Man = 86-92. Caption, triple inter-annotator mAP @ 0.5-0.95 (%).Sci = 94-99. Caption, triple + inter\n---\nf3c18411-85df-4c72-a784-f39f602745ec None\nPage-footer, triple inter-annotator mAP @ 0.5-0.95 (%).Ten + = 96-98. Page-header, Count = 58022. Page-header, % of Total.Train = 5.10. Page-header, % of Total.Test = 6.70. + Page-header, % of Total.Val = 5.06. Page-header, triple inter-annotator mAP @ 0.5-0.95 (%).All = 85-89. Page-header, + triple inter-annotator mAP @ 0.5-0.95 (%).Fin = 66-76. Page-header, triple inter-annotator mAP @ 0.5-0.95 (%).Ma\n---\n6cb167da-a385-4db9-af59-62f23b7cc7b5 + None\nmAP @ 0.5-0.95 (%).Sci = 89-94. Total, triple inter-annotator mAP @ 0.5-0.95 (%).Law = 86-91. Total, triple + inter-annotator mAP @ 0.5-0.95 (%).Pat = 71-76. Total, triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 68-85\nFigure + 3: Corpus Conversion Service annotation user interface. The PDF page is shown in the background, with overlaid text-cells + (in darker shades). The annotation boxes can be drawn \n---\nf647f4ad-7008-40f4-8fe4-f8460da0b5a7 None\n(%).Sci + = 94-100. Title, triple inter-annotator mAP @ 0.5-0.95 (%).Law = 82-96. Title, triple inter-annotator mAP @ 0.5-0.95 + (%).Pat = 68-79. Title, triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 24-56. Total, Count = 1107470. Total, % + of Total.Train = 941123. Total, % of Total.Test = 99816. Total, % of Total.Val = 66531. Total, triple inter-annotator + mAP @ 0.5-0.95 (%).All = 82-83. Total, triple\n---\n70c969da-7627-48d7-9944-5274284c4bdf None\nn/a. Footnote, Count + = 6318. Footnote, % of Total.Train = 0.60. Footnote, % of Total.Test = 0.31. Footnote, % of Total.Val = 0.58. Footnote, + triple inter-annotator mAP @ 0.5-0.95 (%).All = 83-91. Footnote, triple inter-annotator mAP @ 0.5-0.95 (%).Fin = + n/a. Footnote, triple inter-annotator mAP @ 0.5-0.95 (%).Man = 100. Footnote, triple inter-annotator mAP @ 0.5-0.95 + (%).Sci = 62-88. Footnote, tri\n---\n002567d1-7967-4bbb-bf7e-93f6a1bf56d3 None\n0.5-0.95 (%).Law = 87-94. Section-header, + triple inter-annotator mAP @ 0.5-0.95 (%).Pat = 69-73. Section-header, triple inter-annotator mAP @ 0.5-0.95 (%).Ten + = 78-86. Table, Count = 34733. Table, % of Total.Train = 3.20. Table, % of Total.Test = 2.27. Table, % of Total.Val + = 3.60. Table, triple inter-annotator mAP @ 0.5-0.95 (%).All = 77-81. Table, triple inter-annotator mAP @ 0.5-0.95 + (%).Fin = \n---\n6f761413-b87a-436c-b2bb-ec1997478dec None\ninter-annotator mAP @ 0.5-0.95 (%).Pat = 66-71. Picture, + triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 59-76. Section-header, Count = 142884. Section-header, % of Total.Train + = 12.60. Section-header, % of Total.Test = 15.77. Section-header, % of Total.Val = 12.85. Section-header, triple + inter-annotator mAP @ 0.5-0.95 (%).All = 83-84. Section-header, triple inter-annotator mAP @ 0.5-0.95 (%).Fin \n---\neba63d0e-c89f-4ae8-a6ac-4f96788fdb21 + None\n0.5-0.95 (%).Pat = 97-99. Page-header, triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 81-86. Picture, Count + = 45976. Picture, % of Total.Train = 4.21. Picture, % of Total.Test = 2.78. Picture, % of Total.Val = 5.31. Picture, + triple inter-annotator mAP @ 0.5-0.95 (%).All = 69-71. Picture, triple inter-annotator mAP @ 0.5-0.95 (%).Fin = + 56-59. Picture, triple inter-annotator mAP @ 0.5-0.95 (%).Man \n","stderr":"","success":true}' role: tool - tool_call_id: call_jyyqj1qx + tool_call_id: call_4s9httx6 model: gpt-oss reasoning_effort: low stream: false @@ -2485,11 +1291,10 @@ interactions: tools: - function: description: |- - Execute Python code in a Docker-sandboxed environment. + Execute Python code in a sandboxed interpreter. The code has access to haiku.rag functions (search, list_documents, - get_document, get_docling_document, llm) and any Python standard - library module. + get_document, get_chunk, llm). Use print() to output results. @@ -2529,7 +1334,345 @@ interactions: response: headers: content-length: - - '1448' + - '945' + content-type: + - application/json + parsed_body: + choices: + - finish_reason: tool_calls + index: 0 + message: + content: '' + reasoning: 'We see labels: Caption, Footnote, Formula, List-item, Page-footer, Page-header, Picture, Title, Table, + Section-header. Also maybe others. Probably 11 class labels mentioned: Caption, Footnote, Formula, List-item, + Page-footer, Page-header, Picture, Title, Table, Section-header, maybe ? Let''s list unique. Use manual extraction + maybe.' + role: assistant + tool_calls: + - function: + arguments: '{"code":"# extract manually known labels\nlabels=[''Caption'',''Footnote'',''Formula'',''List-item'',''Page-footer'',''Page-header'',''Picture'',''Title'',''Table'',''Section-header'']\nprint(labels)\n"}' + name: execute_code + id: call_2fnlhi2j + index: 0 + type: function + created: 1771924578 + id: chatcmpl-8 + model: gpt-oss + object: chat.completion + system_fingerprint: fp_ollama + usage: + completion_tokens: 138 + prompt_tokens: 4309 + total_tokens: 4447 + status: + code: 200 + message: OK +- request: + headers: + accept: + - application/json + accept-encoding: + - gzip, deflate, zstd + connection: + - keep-alive + content-length: + - '16793' + content-type: + - application/json + host: + - localhost:11434 + method: POST + parsed_body: + messages: + - content: |- + You are a Recursive Language Model (RLM) agent that solves complex research questions by writing and executing Python code. + + IMPORTANT: You MUST use the `execute_code` tool to run Python code. The functions described below are ONLY available inside the execute_code tool - you cannot access them any other way. Always execute code to answer questions; do not just describe what code would do. + + CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just call them with `await`: + - results = await search("query") ✓ CORRECT + - from haiku.rag import search ✗ WRONG - will fail + - results = search("query") ✗ WRONG - must use await + + You have access to a sandboxed Python interpreter with these haiku.rag functions (use them directly with `await`, no imports needed): + + ## Available Functions + + ### await search(query, limit=10) -> list[dict] + Search the knowledge base using hybrid search (vector + full-text). + Returns list of dicts with keys: chunk_id, content, document_id, document_title, document_uri, score, page_numbers, headings + + ### await list_documents(limit=10, offset=0) -> list[dict] + List available documents in the knowledge base. + Returns list of dicts with keys: id, title, uri, created_at + + ### await get_document(id_or_title) -> str | None + Get the full text content of a document by ID, title, or URI. + Returns the document content as a string, or None if not found. + + ### await get_chunk(chunk_id) -> dict | None + Get a specific chunk by its ID (from search results). + Returns dict with keys: chunk_id, content, document_id, document_title, headings, page_numbers, labels + Use this to retrieve full chunk details and metadata for citation. + + ### await llm(prompt) -> str + Call an LLM directly with the given prompt. Returns the response as a string. + Use this for classification, summarization, extraction, or any task where you + already have the content and just need LLM reasoning. + + ## Pre-loaded Documents Variable + + If documents were pre-loaded for this session, a `documents` variable is available: + ```python + # documents is a list of dicts with keys: id, title, uri, content + for doc in documents: + print(doc['title'], len(doc['content'])) + ``` + Check if it exists with: `if 'documents' in dir(): ...` + + ## Available Python Features + + The interpreter supports: variables, arithmetic, strings, f-strings, lists, dicts, tuples, sets, loops, conditionals, comprehensions, functions, async/await, `map()`, `sorted()`/`.sort(key=...)`, try/except, and the `json` module. + + Not supported: imports (other than `json`), class definitions, generators/yield, match statements, decorators, `with` statements. + + For pattern matching or text extraction, use string methods (`str.split`, `str.find`, `str.startswith`, `in` operator) or the `llm()` function. + + ## Strategy Guide + + 1. **Explore First**: Start by listing documents or searching to understand what's available. Document names may differ from filenames (e.g., "tbmed593.pdf" might be stored as "TB MED 593" or similar). + 2. **If get_document returns None**: Use `await list_documents()` to see actual document titles, or `await search()` to find relevant content. + 3. **Iterative Refinement**: Run code, examine results, adjust your approach based on what you find. + 4. **Use print() Liberally**: The sandbox captures stdout - print intermediate results to see what you're working with. + 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and data structures. + 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use `await llm()`. + 7. **Cite Your Sources**: Use get_chunk() to retrieve chunk metadata for citations. Track which documents/chunks informed your answer. + + ## Example Patterns + + ### Counting documents matching a condition + ```python + docs = await list_documents(limit=100) + count = 0 + for doc in docs: + content = await get_document(doc['id']) + if content and 'keyword' in content.lower(): + count += 1 + print(f"Found in: {doc['title']}") + print(f"Total: {count}") + ``` + + ### Extracting data with llm() + ```python + numbers = [] + results = await search("financial data", limit=20) + for r in results: + extracted = await llm(f"Extract all dollar amounts from this text as a comma-separated list of numbers (no $ signs): {r['content']}") + for part in extracted.split(','): + part = part.strip().replace(',', '') + if part.isdigit(): + numbers.append(int(part)) + if numbers: + print(f"Average: {sum(numbers) / len(numbers)}") + ``` + + ### Using search results with get_chunk for citations + ```python + results = await search("safety requirements", limit=5) + for r in results: + chunk = await get_chunk(r['chunk_id']) + print(f"From '{chunk['document_title']}', page {chunk['page_numbers']}: {chunk['content'][:100]}") + ``` + + ### Using llm() for classification + ```python + content = await get_document("Q1 Report") + sentiment = await llm(f"Classify the sentiment as positive, negative, or mixed: {content}") + print(sentiment) + ``` + + ## Workflow + + 1. **ALWAYS start by using execute_code** to explore the knowledge base + 2. Run multiple code blocks as needed to gather information + 3. After collecting data, provide your final answer + + ## Output Format + + CRITICAL: Your final response MUST be valid JSON matching this exact schema: + ```json + {"answer": "Your complete answer here as a string", "program": "Your final consolidated program here as a string"} + ``` + + - `answer`: A clear answer to the user's question with key findings and references to specific documents/chunks. + - `program`: A single, self-contained Python program that produces the answer. Consolidate your exploratory code executions into one clean script. + + Do NOT return arbitrary JSON structures. Always use the exact format: {"answer": "...", "program": "..."} + + CRITICAL: You MUST call execute_code at least once before providing your answer. Never give up without trying to execute code first. + role: system + - content: Search for content about document element types or labels. What are all the different document element types + mentioned? List them all. + role: user + - content: null + reasoning: Need to search for document element types or labels. Use search. + role: assistant + tool_calls: + - function: + arguments: '{"code":"results = await search(\"document element types\" , limit=20)\nprint(len(results))\nfor r + in results[:5]:\n print(r[''chunk_id''], r[''document_title''], r[''content''][:200])\n"}' + name: execute_code + id: call_j7uw3i6n + type: function + - content: '{"code":"results = await search(\"document element types\" , limit=20)\nprint(len(results))\nfor r in results[:5]:\n print(r[''chunk_id''], + r[''document_title''], r[''content''][:200])\n","stdout":"17\n7c3a912c-a769-49e9-a661-402d7d64ec9f None Phase 2: + Label selection and guideline. We reviewed the collected documents and identified the most common structural features + they exhibit. This was achieved by identifying recurrent layout elements \n62e98936-a694-4cb0-a1a4-409b0121e08d + None Phase 1: Data selection and preparation. Our inclusion criteria for documents were described in Section 3. + A large effort went into ensuring that all documents are free to use. The data sources includ\nd583f0ed-c230-4f8e-988a-86875c0c633d + None $_{Affiliation}$, as seen in DocBank, are often only distinguishable by discriminating on\nPreparation work + included uploading and parsing the sourced PDF documents in the Corpus Conversion Service (CC\n180df00a-2a5b-4e2a-bfcc-4cd67fbc9db0 + None \nTable 1: DocLayNet dataset overview. Along with the frequency of each class label, we present the relative + occurrence (as % of row \"Total\") in the train, test and validation sets. The inter-annotator\nf3c18411-85df-4c72-a784-f39f602745ec + None Page-footer, triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 96-98. Page-header, Count = 58022. Page-header, + % of Total.Train = 5.10. Page-header, % of Total.Test = 6.70. Page-header, % of Total.Val =\n","stderr":"","success":true}' + role: tool + tool_call_id: call_j7uw3i6n + - content: null + reasoning: We need list all document element types or labels mentioned. Likely these are labels in documents like + "Page-header", "Page-footer", etc. Let's search more. + role: assistant + tool_calls: + - function: + arguments: '{"code":"results = await search(\"label\" , limit=30)\nlabels=set()\nfor r in results:\n # look + for uppercase words separated by hyphens\n import re\n for m in re.findall(r\"([A-Z][A-Za-z]+(?:-[A-Z][A-Za-z]+)*)\", + r[''content'']):\n labels.add(m)\nprint(sorted(labels))\n"}' + name: execute_code + id: call_gzg850pl + type: function + - content: '{"code":"results = await search(\"label\" , limit=30)\nlabels=set()\nfor r in results:\n # look for uppercase + words separated by hyphens\n import re\n for m in re.findall(r\"([A-Z][A-Za-z]+(?:-[A-Z][A-Za-z]+)*)\", r[''content'']):\n labels.add(m)\nprint(sorted(labels))\n","stdout":"","stderr":"ModuleNotFoundError: + No module named ''re''","success":false}' + role: tool + tool_call_id: call_gzg850pl + - content: null + reasoning: Import not allowed. Use string methods. Probably easier to manually inspect. Let's list results. + role: assistant + tool_calls: + - function: + arguments: '{"code":"results = await search(\"label\" , limit=10)\nfor r in results:\n print(''---'')\n print(r[''chunk_id''], + r[''document_title''])\n print(r[''content''][:400])\n"}' + name: execute_code + id: call_4s9httx6 + type: function + - content: '{"code":"results = await search(\"label\" , limit=10)\nfor r in results:\n print(''---'')\n print(r[''chunk_id''], + r[''document_title''])\n print(r[''content''][:400])\n","stdout":"---\n7c3a912c-a769-49e9-a661-402d7d64ec9f None\nPhase + 2: Label selection and guideline. We reviewed the collected documents and identified the most common structural + features they exhibit. This was achieved by identifying recurrent layout elements and lead us to the definition + of 11 distinct class labels. These 11 class labels are $_{Caption}$, $_{Footnote}$, $_{Formula}$, $_{List-item}$, + Page-$_{footer}$, $_{Page-header}$, $_{Picture}$, $_{Sec\n---\n180df00a-2a5b-4e2a-bfcc-4cd67fbc9db0 None\n\nTable + 1: DocLayNet dataset overview. Along with the frequency of each class label, we present the relative occurrence + (as % of row \"Total\") in the train, test and validation sets. The inter-annotator agreement is computed as the + mAP@0.5-0.95 metric between pairwise annotations from the triple-annotated pages, from which we obtain accuracy + ranges.\n---\n5c6655e2-40d8-4ea8-9741-de6a98895449 None\nCaption, Count = 22524. Caption, % of Total.Train = 2.04. + Caption, % of Total.Test = 1.77. Caption, % of Total.Val = 2.32. Caption, triple inter-annotator mAP @ 0.5-0.95 + (%).All = 84-89. Caption, triple inter-annotator mAP @ 0.5-0.95 (%).Fin = 40-61. Caption, triple inter-annotator + mAP @ 0.5-0.95 (%).Man = 86-92. Caption, triple inter-annotator mAP @ 0.5-0.95 (%).Sci = 94-99. Caption, triple + inter\n---\nf3c18411-85df-4c72-a784-f39f602745ec None\nPage-footer, triple inter-annotator mAP @ 0.5-0.95 (%).Ten + = 96-98. Page-header, Count = 58022. Page-header, % of Total.Train = 5.10. Page-header, % of Total.Test = 6.70. + Page-header, % of Total.Val = 5.06. Page-header, triple inter-annotator mAP @ 0.5-0.95 (%).All = 85-89. Page-header, + triple inter-annotator mAP @ 0.5-0.95 (%).Fin = 66-76. Page-header, triple inter-annotator mAP @ 0.5-0.95 (%).Ma\n---\n6cb167da-a385-4db9-af59-62f23b7cc7b5 + None\nmAP @ 0.5-0.95 (%).Sci = 89-94. Total, triple inter-annotator mAP @ 0.5-0.95 (%).Law = 86-91. Total, triple + inter-annotator mAP @ 0.5-0.95 (%).Pat = 71-76. Total, triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 68-85\nFigure + 3: Corpus Conversion Service annotation user interface. The PDF page is shown in the background, with overlaid text-cells + (in darker shades). The annotation boxes can be drawn \n---\nf647f4ad-7008-40f4-8fe4-f8460da0b5a7 None\n(%).Sci + = 94-100. Title, triple inter-annotator mAP @ 0.5-0.95 (%).Law = 82-96. Title, triple inter-annotator mAP @ 0.5-0.95 + (%).Pat = 68-79. Title, triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 24-56. Total, Count = 1107470. Total, % + of Total.Train = 941123. Total, % of Total.Test = 99816. Total, % of Total.Val = 66531. Total, triple inter-annotator + mAP @ 0.5-0.95 (%).All = 82-83. Total, triple\n---\n70c969da-7627-48d7-9944-5274284c4bdf None\nn/a. Footnote, Count + = 6318. Footnote, % of Total.Train = 0.60. Footnote, % of Total.Test = 0.31. Footnote, % of Total.Val = 0.58. Footnote, + triple inter-annotator mAP @ 0.5-0.95 (%).All = 83-91. Footnote, triple inter-annotator mAP @ 0.5-0.95 (%).Fin = + n/a. Footnote, triple inter-annotator mAP @ 0.5-0.95 (%).Man = 100. Footnote, triple inter-annotator mAP @ 0.5-0.95 + (%).Sci = 62-88. Footnote, tri\n---\n002567d1-7967-4bbb-bf7e-93f6a1bf56d3 None\n0.5-0.95 (%).Law = 87-94. Section-header, + triple inter-annotator mAP @ 0.5-0.95 (%).Pat = 69-73. Section-header, triple inter-annotator mAP @ 0.5-0.95 (%).Ten + = 78-86. Table, Count = 34733. Table, % of Total.Train = 3.20. Table, % of Total.Test = 2.27. Table, % of Total.Val + = 3.60. Table, triple inter-annotator mAP @ 0.5-0.95 (%).All = 77-81. Table, triple inter-annotator mAP @ 0.5-0.95 + (%).Fin = \n---\n6f761413-b87a-436c-b2bb-ec1997478dec None\ninter-annotator mAP @ 0.5-0.95 (%).Pat = 66-71. Picture, + triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 59-76. Section-header, Count = 142884. Section-header, % of Total.Train + = 12.60. Section-header, % of Total.Test = 15.77. Section-header, % of Total.Val = 12.85. Section-header, triple + inter-annotator mAP @ 0.5-0.95 (%).All = 83-84. Section-header, triple inter-annotator mAP @ 0.5-0.95 (%).Fin \n---\neba63d0e-c89f-4ae8-a6ac-4f96788fdb21 + None\n0.5-0.95 (%).Pat = 97-99. Page-header, triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 81-86. Picture, Count + = 45976. Picture, % of Total.Train = 4.21. Picture, % of Total.Test = 2.78. Picture, % of Total.Val = 5.31. Picture, + triple inter-annotator mAP @ 0.5-0.95 (%).All = 69-71. Picture, triple inter-annotator mAP @ 0.5-0.95 (%).Fin = + 56-59. Picture, triple inter-annotator mAP @ 0.5-0.95 (%).Man \n","stderr":"","success":true}' + role: tool + tool_call_id: call_4s9httx6 + - content: null + reasoning: 'We see labels: Caption, Footnote, Formula, List-item, Page-footer, Page-header, Picture, Title, Table, + Section-header. Also maybe others. Probably 11 class labels mentioned: Caption, Footnote, Formula, List-item, Page-footer, + Page-header, Picture, Title, Table, Section-header, maybe ? Let''s list unique. Use manual extraction maybe.' + role: assistant + tool_calls: + - function: + arguments: '{"code":"# extract manually known labels\nlabels=[''Caption'',''Footnote'',''Formula'',''List-item'',''Page-footer'',''Page-header'',''Picture'',''Title'',''Table'',''Section-header'']\nprint(labels)\n"}' + name: execute_code + id: call_2fnlhi2j + type: function + - content: '{"code":"# extract manually known labels\nlabels=[''Caption'',''Footnote'',''Formula'',''List-item'',''Page-footer'',''Page-header'',''Picture'',''Title'',''Table'',''Section-header'']\nprint(labels)\n","stdout":"[''Caption'', + ''Footnote'', ''Formula'', ''List-item'', ''Page-footer'', ''Page-header'', ''Picture'', ''Title'', ''Table'', ''Section-header'']\n","stderr":"","success":true}' + role: tool + tool_call_id: call_2fnlhi2j + model: gpt-oss + reasoning_effort: low + stream: false + tool_choice: auto + tools: + - function: + description: |- + Execute Python code in a sandboxed interpreter. + + The code has access to haiku.rag functions (search, list_documents, + get_document, get_chunk, llm). + + Use print() to output results. + + Structured result with success status, stdout, and stderr. + + name: execute_code + parameters: + additionalProperties: false + properties: + code: + description: Python code to execute. + type: string + required: + - code + type: object + strict: true + type: function + - function: + description: Result from RLM agent execution. + name: final_result + parameters: + additionalProperties: false + properties: + answer: + description: The answer to the user's question + type: string + program: + description: The final consolidated program + type: string + required: + - answer + - program + type: object + strict: true + type: function + uri: http://localhost:11434/v1/chat/completions + response: + headers: + content-length: + - '1797' content-type: - application/json parsed_body: @@ -2537,491 +1680,28 @@ interactions: - finish_reason: stop index: 0 message: - content: "**Answer**\n\nThe DocLayNet dataset defines the following 11 document element types (class labels):\n\n - \ - Caption \n - Footnote \n - Formula \n - List‑item \n - Page‑footer \n - Page‑header \n - Picture - \ \n - Section‑header \n - Table \n - Text \n - Title \n\nThese are all the distinct element types mentioned - in the repository.\n\n**Program**\n\n```python\n# Find the DocLayNet table in the knowledge base and extract the - 11 labels\nimport re\n\n# Search for the table that lists class labels\nresults = search('Table 1: DocLayNet dataset - overview', limit=5)\n\nlabels = set()\nfor r in results:\n # Look for the part that lists class labels in the - content\n m = re.search(r'lead us to the definition of ([^.]*)\\. These ([^.]*)', r['content'])\n if m:\n - \ # Split by comma and strip whitespace and symbols\n parts = re.split(r',\\s*', m.group(2))\n for - p in parts:\n p = re.sub(r'^[\\\\$\\\\{\\\\}\\\\s]*|[\\\\$\\\\{\\\\}\\\\s]*$', '', p)\n if - p:\n labels.add(p)\n \nprint(\"Found labels:\")\nfor lbl in sorted(labels):\n print(f\"- - {lbl}\")\n```" + content: '{"answer":"The documents describe a total of **10 distinct document element types (labels)**:\n\n1. **Caption**\n2. + **Footnote**\n3. **Formula**\n4. **List-item**\n5. **Page‑header**\n6. **Page‑footer**\n7. **Picture**\n8. **Title**\n9. + **Table**\n10. **Section‑header**","program":"# Consolidated script to discover document element types\nresults + = await search(\"document element types\", limit=50)\nlabels = set()\nfor r in results:\n # Search for capitalised + words that match the known classes\n # This uses simple string checks because imports are not allowed.\n if + \"Caption\" in r[''content'']:\n labels.add(\"Caption\")\n if \"Footnote\" in r[''content'']:\n labels.add(\"Footnote\")\n if + \"Formula\" in r[''content'']:\n labels.add(\"Formula\")\n if \"List-item\" in r[''content'']:\n labels.add(\"List-item\")\n if + \"Page-header\" in r[''content'']:\n labels.add(\"Page-header\")\n if \"Page-footer\" in r[''content'']:\n labels.add(\"Page-footer\")\n if + \"Picture\" in r[''content'']:\n labels.add(\"Picture\")\n if \"Title\" in r[''content'']:\n labels.add(\"Title\")\n if + \"Table\" in r[''content'']:\n labels.add(\"Table\")\n if \"Section-header\" in r[''content'']:\n labels.add(\"Section-header\")\nprint(\"Found + labels:\", sorted(labels))"}' + reasoning: Answer list. role: assistant - created: 1770373449 - id: chatcmpl-965 + created: 1771924587 + id: chatcmpl-392 model: gpt-oss object: chat.completion system_fingerprint: fp_ollama usage: - completion_tokens: 291 - prompt_tokens: 5962 - total_tokens: 6253 - status: - code: 200 - message: OK -- request: - headers: - accept: - - application/json - accept-encoding: - - gzip, deflate, zstd - connection: - - keep-alive - content-length: - - '23396' - content-type: - - application/json - host: - - localhost:11434 - method: POST - parsed_body: - messages: - - content: |- - You are a Recursive Language Model (RLM) agent that solves complex research questions by writing and executing Python code. - - IMPORTANT: You MUST use the `execute_code` tool to run Python code. The functions described below are ONLY available inside the execute_code tool - you cannot access them any other way. Always execute code to answer questions; do not just describe what code would do. - - CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just use them directly: - - search("query") ✓ CORRECT - - from haiku.rag import search ✗ WRONG - will fail - - You have access to a sandboxed Python environment with these haiku.rag functions (use them directly, no imports needed): - - ## Available Functions - - ### search(query, limit=10) -> list[dict] - Search the knowledge base using hybrid search (vector + full-text). - Returns list of dicts with keys: chunk_id, content, document_id, document_title, document_uri, score, page_numbers, headings - - ### list_documents(limit=10, offset=0) -> list[dict] - List available documents in the knowledge base. - Returns list of dicts with keys: id, title, uri, created_at - - ### get_document(id_or_title) -> str | None - Get the full text content of a document by ID, title, or URI. - Returns the document content as a string, or None if not found. - - ### get_docling_document(id_or_title) -> DoclingDocument | None - Get the structured DoclingDocument object for advanced analysis. - Returns a DoclingDocument object, or None if not found. - See "DoclingDocument API" section below for how to use it. - - ### llm(prompt) -> str - Call an LLM directly with the given prompt. Returns the response as a string. - Use this for classification, summarization, extraction, or any task where you - already have the content and just need LLM reasoning. - - ## Pre-loaded Documents Variable - - If documents were pre-loaded for this session, a `documents` variable is available: - ```python - # documents is a list of dicts with keys: id, title, uri, content - for doc in documents: - print(doc['title'], len(doc['content'])) - ``` - Check if it exists with: `if 'documents' in dir(): ...` - - ## Standard Library Modules - You can import any Python standard library module. - - ## Strategy Guide - - 1. **Explore First**: Start by listing documents or searching to understand what's available. Document names may differ from filenames (e.g., "tbmed593.pdf" might be stored as "TB MED 593" or similar). - 2. **If get_document returns None**: Use `list_documents()` to see actual document titles, or `search()` to find relevant content. - 3. **Iterative Refinement**: Run code, examine results, adjust your approach based on what you find. - 4. **Use print() Liberally**: The REPL captures stdout - print intermediate results to see what you're working with. - 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and use collections. - 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use llm(). - 7. **Cite Your Sources**: Track which documents/chunks informed your answer for citation. - - ## DoclingDocument API - - When you call `get_docling_document(id_or_title)`, you get a DoclingDocument object for structured document analysis. - - ### Properties - - `doc.texts` - List of all text items (paragraphs, headings, etc.) - - `doc.tables` - List of all tables - - `doc.pictures` - List of all pictures/figures - - `doc.name` - Document name - - ### Methods - - `doc.iterate_items(with_groups=False)` - Iterate all items with hierarchy level - Returns tuples of (item, level) where level is nesting depth - - `doc.export_to_markdown()` - Export entire document as markdown string - - ### Text Item Properties - - `item.text` - The text content - - `item.label` - Type: title, paragraph, section_header, list_item, etc. (lowercase enum values) - - `item.prov` - Provenance (page numbers, bounding boxes) - - ### Table Access - - `table.data.num_rows`, `table.data.num_cols` - Dimensions - - `table.data.table_cells` - List of TableCell objects - - `cell.text`, `cell.start_row_offset_idx`, `cell.start_col_offset_idx` - - ### Example Usage - ```python - doc = get_docling_document("My Document") - - # Get all headings - headings = [t.text for t in doc.texts if "header" in str(t.label)] - - # Iterate with structure - for item, level in doc.iterate_items(): - print(" " * level + item.text[:50]) - - # Extract table data - for table in doc.tables: - for cell in table.data.table_cells: - print(f"Row {cell.start_row_offset_idx}, Col {cell.start_col_offset_idx}: {cell.text}") - ``` - - ## Example Patterns - - ### Counting documents matching a condition - ```python - docs = list_documents(limit=100) - count = 0 - for doc in docs: - content = get_document(doc['id']) - if content and 'keyword' in content.lower(): - count += 1 - print(f"Found in: {doc['title']}") - print(f"Total: {count}") - ``` - - ### Aggregating data across documents - ```python - import re - numbers = [] - results = search("financial data", limit=20) - for r in results: - matches = re.findall(r'\$([\d,]+)', r['content']) - for m in matches: - numbers.append(int(m.replace(',', ''))) - print(f"Average: ${sum(numbers)/len(numbers):,.2f}") - ``` - - ### Using llm() for classification - ```python - # Get document content - content = get_document("Q1 Report") - # Use llm() to classify sentiment - sentiment = llm(f"Classify the sentiment as positive, negative, or mixed: {content}") - print(sentiment) - ``` - - ## Workflow - - 1. **ALWAYS start by using execute_code** to explore the knowledge base - 2. Run multiple code blocks as needed to gather information - 3. After collecting data, provide your final answer - - ## Output Format - - CRITICAL: Your final response MUST be valid JSON matching this exact schema: - ```json - {"answer": "Your complete answer here as a string", "program": "Your final consolidated program here as a string"} - ``` - - - `answer`: A clear answer to the user's question with key findings and references to specific documents/chunks. - - `program`: A single, self-contained Python program that produces the answer. Consolidate your exploratory code executions into one clean script. - - Do NOT return arbitrary JSON structures. Always use the exact format: {"answer": "...", "program": "..."} - - CRITICAL: You MUST call execute_code at least once before providing your answer. Never give up without trying to execute code first. - role: system - - content: Search for content about document element types or labels. What are all the different document element types - mentioned? List them all. - role: user - - content: |- - - We need search for "document element types" or "labels". We'll run search. - - role: assistant - tool_calls: - - function: - arguments: '{"code":"results = search(''document element types labels'', limit=20)\nprint(len(results))\nfor r - in results:\n print(r[''document_title''], r[''chunk_id''], r[''score''])\n print(r[''content''][:200])\n print(''---'')"}' - name: execute_code - id: call_o2vpon36 - type: function - - content: '{"code":"results = search(''document element types labels'', limit=20)\nprint(len(results))\nfor r in results:\n print(r[''document_title''], - r[''chunk_id''], r[''score''])\n print(r[''content''][:200])\n print(''---'')","stdout":"17\nNone 360a55ef-1fae-45c7-9506-d681bf7d2642 - 0.032786883413791656\nPhase 2: Label selection and guideline. We reviewed the collected documents and identified - the most common structural features they exhibit. This was achieved by identifying recurrent layout elements \n---\nNone - f71a4730-2d9c-4317-9e33-2488adda8f3d 0.0317540317773819\n\nTable 1: DocLayNet dataset overview. Along with the frequency - of each class label, we present the relative occurrence (as % of row \"Total\") in the train, test and validation - sets. The inter-annotator\n---\nNone 994f8aeb-bdf3-434d-9b2e-69d4a6a9c623 0.03015873022377491\n$_{Affiliation}$, - as seen in DocBank, are often only distinguishable by discriminating on\nPreparation work included uploading and - parsing the sourced PDF documents in the Corpus Conversion Service (CC\n---\nNone 70a1c951-bc93-4302-95ca-bdbb832f3cf9 - 0.029462365433573723\nPhase 1: Data selection and preparation. Our inclusion criteria for documents were described - in Section 3. A large effort went into ensuring that all documents are free to use. The data sources includ\n---\nNone - 916ed8c5-d868-4064-a459-1f2cc704df4e 0.028371628373861313\nmAP @ 0.5-0.95 (%).Sci = 89-94. Total, triple inter-annotator - mAP @ 0.5-0.95 (%).Law = 86-91. Total, triple inter-annotator mAP @ 0.5-0.95 (%).Pat = 71-76. Total, triple inter-annotator - mAP @ 0.5-0.95\n---\nNone 41a5b3f5-ff96-4856-9eb9-4695fe28b39c 0.01587301678955555\nCaption, Count = 22524. Caption, - % of Total.Train = 2.04. Caption, % of Total.Test = 1.77. Caption, % of Total.Val = 2.32. Caption, triple inter-annotator - mAP @ 0.5-0.95 (%).All = 84-89. Caption, trip\n---\nNone c359d67a-0809-45bb-bfb0-139817b967fd 0.015625\nPage-footer, - triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 96-98. Page-header, Count = 58022. Page-header, % of Total.Train - = 5.10. Page-header, % of Total.Test = 6.70. Page-header, % of Total.Val =\n---\nNone 50e95bc9-862e-4bdb-9e8c-4d1e38d0eee6 - 0.015384615398943424\n0.5-0.95 (%).Law = 87-94. Section-header, triple inter-annotator mAP @ 0.5-0.95 (%).Pat = - 69-73. Section-header, triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 78-86. Table, Count = 34733. Table, % o\n---\nNone - d40d3add-cd91-4774-9d3e-77c388e0f9a4 0.01515151560306549\n89-93. Text, triple inter-annotator mAP @ 0.5-0.95 (%).Law - = 87-92. Text, triple inter-annotator mAP @ 0.5-0.95 (%).Pat = 71-79. Text, triple inter-annotator mAP @ 0.5-0.95 - (%).Ten = 87-95. Title, Cou\n---\nNone 683d9ab8-6363-4165-9cc7-b89145db3f33 0.014925372786819935\n185660. List-item, - % of Total.Train = 17.19. List-item, % of Total.Test = 13.34. List-item, % of Total.Val = 15.82. List-item, triple - inter-annotator mAP @ 0.5-0.95 (%).All = 87-88. List-item, triple \n---\nNone f5d1d638-002e-42ab-8c35-6fbf834ab435 - 0.014705882407724857\ninter-annotator mAP @ 0.5-0.95 (%).Pat = 66-71. Picture, triple inter-annotator mAP @ 0.5-0.95 - (%).Ten = 59-76. Section-header, Count = 142884. Section-header, % of Total.Train = 12.60. Section-header\n---\nNone - 171eb4b0-e518-4e65-9ef4-5655789dceae 0.014492753893136978\nn/a. Footnote, Count = 6318. Footnote, % of Total.Train - = 0.60. Footnote, % of Total.Test = 0.31. Footnote, % of Total.Val = 0.58. Footnote, triple inter-annotator mAP - @ 0.5-0.95 (%).All = 83-91. Foot\n---\nNone 64741ed7-d686-4e6c-b8e1-462a51282ecc 0.014084506779909134\n(%).Sci = - 94-100. Title, triple inter-annotator mAP @ 0.5-0.95 (%).Law = 82-96. Title, triple inter-annotator mAP @ 0.5-0.95 - (%).Pat = 68-79. Title, triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 24-\n---\nNone d8771358-ce2e-407f-af30-ee410ed8a579 - 0.013888888992369175\n0.5-0.95 (%).Pat = 97-99. Page-header, triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 81-86. - Picture, Count = 45976. Picture, % of Total.Train = 4.21. Picture, % of Total.Test = 2.78. Picture, % of T\n---\nNone - c2289d48-99bf-4400-ac39-155b764485db 0.013698630034923553\ninter-annotator mAP @ 0.5-0.95 (%).Law = 58-80. Table, - triple inter-annotator mAP @ 0.5-0.95 (%).Pat = 79-84. Table, triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 70-85. - Text, Count = 510377. Text, \n---\nNone 6a21d905-7202-4ed9-9241-32934c09f36e 0.013513513840734959\n0.5-0.95 (%).Ten - = 93-95. Page-footer, Count = 70878. Page-footer, % of Total.Train = 6.51. Page-footer, % of Total.Test = 5.58. - Page-footer, % of Total.Val = 6.00. Page-footer, triple inter-annotator\n---\nNone 610e94fb-1051-4e09-b94b-98cd485e551f - 0.01315789483487606\n= 82-97. Formula, Count = 25027. Formula, % of Total.Train = 2.25. Formula, % of Total.Test - = 1.90. Formula, % of Total.Val = 2.96. Formula, triple inter-annotator mAP @ 0.5-0.95 (%).All = 83-85. Form\n---\n","stderr":"","success":true}' - role: tool - tool_call_id: call_o2vpon36 - - content: |- - - Chapters mention labels like "page-header", "page-footer", "section-header", "picture", "table", "text", "title", "list-item", "footnote", "formula". Also maybe "Caption". We'll extract unique set. - - role: assistant - tool_calls: - - function: - arguments: '{"code":"labels=set()\nfor r in results:\n if ''label'' in r[''content'']:\n pass\n","stdout":"","stderr":"FileNotFoundError: - [Errno 2] No such file or directory: ''results''\n","success":false}' - name: execute_code - id: call_1mr5hrfc - type: function - - content: |- - 3 validation errors: - ```json - [ - { - "type": "extra_forbidden", - "loc": [ - "stdout" - ], - "msg": "Extra inputs are not permitted", - "input": "" - }, - { - "type": "extra_forbidden", - "loc": [ - "stderr" - ], - "msg": "Extra inputs are not permitted", - "input": "FileNotFoundError: [Errno 2] No such file or directory: 'results'\n" - }, - { - "type": "extra_forbidden", - "loc": [ - "success" - ], - "msg": "Extra inputs are not permitted", - "input": false - } - ] - ``` - - Fix the errors and try again. - role: tool - tool_call_id: call_1mr5hrfc - - content: |- - - Need to store results variable outside. Let's re-run search and store. - - role: assistant - tool_calls: - - function: - arguments: '{"code":"results = search(''document element types labels'', limit=20)\nlabels=set()\nimport re\nfor - r in results:\n # extract words like \"Caption\", \"Formula\", etc by regex\n for m in re.finditer(r\"\\b([A-Za-z-]+)\\b\", - r[''content'']):\n labels.add(m.group(1))\nprint(sorted(labels)[:30])\nprint(''count'', len(labels))"}' - name: execute_code - id: call_u9fkvji9 - type: function - - content: '{"code":"results = search(''document element types labels'', limit=20)\nlabels=set()\nimport re\nfor r in - results:\n # extract words like \"Caption\", \"Formula\", etc by regex\n for m in re.finditer(r\"\\b([A-Za-z-]+)\\b\", - r[''content'']):\n labels.add(m.group(1))\nprint(sorted(labels)[:30])\nprint(''count'', len(labels))","stdout":"[''-'', - ''A'', ''Abstract'', ''Affiliation'', ''All'', ''Along'', ''Articles'', ''Author'', ''CCS'', ''Caption'', ''Conversion'', - ''Corpus'', ''Count'', ''Critical'', ''Data'', ''DocBank'', ''DocLayNet'', ''Figure'', ''Fin'', ''Footnote'', ''For'', - ''Formula'', ''Label'', ''Labels'', ''Law'', ''List-item'', ''Man'', ''Our'', ''PDF'', ''Page'']\ncount 295\n","stderr":"","success":true}' - role: tool - tool_call_id: call_u9fkvji9 - - content: |- - - Too generic. Instead, use label list from dataset. Let's search for "label\" in dataset info. - - role: assistant - tool_calls: - - function: - arguments: '{"code":"results = search(''label\\n'', limit=20)\nprint(len(results))\nfor r in results[:3]:\n print(r[''document_title''], - r[''chunk_id''], r[''score''])\n print(r[''content''][:200])\n print(''---'')"}' - name: execute_code - id: call_8m1hy7q3 - type: function - - content: '{"code":"results = search(''label\\n'', limit=20)\nprint(len(results))\nfor r in results[:3]:\n print(r[''document_title''], - r[''chunk_id''], r[''score''])\n print(r[''content''][:200])\n print(''---'')","stdout":"17\nNone 360a55ef-1fae-45c7-9506-d681bf7d2642 - 0.032786883413791656\nPhase 2: Label selection and guideline. We reviewed the collected documents and identified - the most common structural features they exhibit. This was achieved by identifying recurrent layout elements \n---\nNone - f71a4730-2d9c-4317-9e33-2488adda8f3d 0.03151364624500275\n\nTable 1: DocLayNet dataset overview. Along with the - frequency of each class label, we present the relative occurrence (as % of row \"Total\") in the train, test and - validation sets. The inter-annotator\n---\nNone 916ed8c5-d868-4064-a459-1f2cc704df4e 0.02903090976178646\nmAP @ - 0.5-0.95 (%).Sci = 89-94. Total, triple inter-annotator mAP @ 0.5-0.95 (%).Law = 86-91. Total, triple inter-annotator - mAP @ 0.5-0.95 (%).Pat = 71-76. Total, triple inter-annotator mAP @ 0.5-0.95\n---\n","stderr":"","success":true}' - role: tool - tool_call_id: call_8m1hy7q3 - - content: |- - - Better: list of element types from dataset: title, heading, paragraph, list_item, table, picture, formula, footnote, caption, page_header, page_footer, section_header. Let's extract from dataset table. Search for "Table 1:" maybe content lists types. - - role: assistant - tool_calls: - - function: - arguments: '{"code":"results = search(''Table 1: DocLayNet'', limit=5)\nfor r in results:\n print(r[''content''][:500])"}' - name: execute_code - id: call_jyyqj1qx - type: function - - content: '{"code":"results = search(''Table 1: DocLayNet'', limit=5)\nfor r in results:\n print(r[''content''][:500])","stdout":"\nTable - 1: DocLayNet dataset overview. Along with the frequency of each class label, we present the relative occurrence - (as % of row \"Total\") in the train, test and validation sets. The inter-annotator agreement is computed as the - mAP@0.5-0.95 metric between pairwise annotations from the triple-annotated pages, from which we obtain accuracy - ranges.\nPhase 2: Label selection and guideline. We reviewed the collected documents and identified the most common - structural features they exhibit. This was achieved by identifying recurrent layout elements and lead us to the - definition of 11 distinct class labels. These 11 class labels are $_{Caption}$, $_{Footnote}$, $_{Formula}$, $_{List-item}$, - Page-$_{footer}$, $_{Page-header}$, $_{Picture}$, $_{Section-header}$, $_{Table}$, $_{Text}$, and $_{Title}$. Critical - factors that were considered for the \n0.5-0.95 (%).Law = 87-94. Section-header, triple inter-annotator mAP @ 0.5-0.95 - (%).Pat = 69-73. Section-header, triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 78-86. Table, Count = 34733. Table, - % of Total.Train = 3.20. Table, % of Total.Test = 2.27. Table, % of Total.Val = 3.60. Table, triple inter-annotator - mAP @ 0.5-0.95 (%).All = 77-81. Table, triple inter-annotator mAP @ 0.5-0.95 (%).Fin = 75-80. Table, triple inter-annotator - mAP @ 0.5-0.95 (%).Man = 83-86. Table, triple inter-annotator m\ninter-annotator mAP @ 0.5-0.95 (%).Law = 58-80. - Table, triple inter-annotator mAP @ 0.5-0.95 (%).Pat = 79-84. Table, triple inter-annotator mAP @ 0.5-0.95 (%).Ten - = 70-85. Text, Count = 510377. Text, % of Total.Train = 45.82. Text, % of Total.Test = 49.28. Text, % of Total.Val - = 45.00. Text, triple inter-annotator mAP @ 0.5-0.95 (%).All = 84-86. Text, triple inter-annotator mAP @ 0.5-0.95 - (%).Fin = 81-86. Text, triple inter-annotator mAP @ 0.5-0.95 (%).Man = 88-93. Text, triple inter-annotator m\nPage-footer, - triple inter-annotator mAP @ 0.5-0.95 (%).Ten = 96-98. Page-header, Count = 58022. Page-header, % of Total.Train - = 5.10. Page-header, % of Total.Test = 6.70. Page-header, % of Total.Val = 5.06. Page-header, triple inter-annotator - mAP @ 0.5-0.95 (%).All = 85-89. Page-header, triple inter-annotator mAP @ 0.5-0.95 (%).Fin = 66-76. Page-header, - triple inter-annotator mAP @ 0.5-0.95 (%).Man = 90-94. Page-header, triple inter-annotator mAP @ 0.5-0.95 (%).Sci - = 98-100. Page-header, triple \n","stderr":"","success":true}' - role: tool - tool_call_id: call_jyyqj1qx - - content: "**Answer**\n\nThe DocLayNet dataset defines the following 11 document element types (class labels):\n\n - \ - Caption \n - Footnote \n - Formula \n - List‑item \n - Page‑footer \n - Page‑header \n - Picture - \ \n - Section‑header \n - Table \n - Text \n - Title \n\nThese are all the distinct element types mentioned - in the repository.\n\n**Program**\n\n```python\n# Find the DocLayNet table in the knowledge base and extract the - 11 labels\nimport re\n\n# Search for the table that lists class labels\nresults = search('Table 1: DocLayNet dataset - overview', limit=5)\n\nlabels = set()\nfor r in results:\n # Look for the part that lists class labels in the - content\n m = re.search(r'lead us to the definition of ([^.]*)\\. These ([^.]*)', r['content'])\n if m:\n - \ # Split by comma and strip whitespace and symbols\n parts = re.split(r',\\s*', m.group(2))\n for - p in parts:\n p = re.sub(r'^[\\\\$\\\\{\\\\}\\\\s]*|[\\\\$\\\\{\\\\}\\\\s]*$', '', p)\n if - p:\n labels.add(p)\n \nprint(\"Found labels:\")\nfor lbl in sorted(labels):\n print(f\"- - {lbl}\")\n```" - role: assistant - - content: |- - 1 validation error: - ```json - [ - { - "type": "json_invalid", - "loc": [], - "msg": "Invalid JSON: expected value at line 1 column 1", - "input": "**Answer**\n\nThe DocLayNet dataset defines the following 11 document element types (class labels):\n\n - Caption \n - Footnote \n - Formula \n - List‑item \n - Page‑footer \n - Page‑header \n - Picture \n - Section‑header \n - Table \n - Text \n - Title \n\nThese are all the distinct element types mentioned in the repository.\n\n**Program**\n\n```python\n# Find the DocLayNet table in the knowledge base and extract the 11 labels\nimport re\n\n# Search for the table that lists class labels\nresults = search('Table 1: DocLayNet dataset overview', limit=5)\n\nlabels = set()\nfor r in results:\n # Look for the part that lists class labels in the content\n m = re.search(r'lead us to the definition of ([^.]*)\\. These ([^.]*)', r['content'])\n if m:\n # Split by comma and strip whitespace and symbols\n parts = re.split(r',\\s*', m.group(2))\n for p in parts:\n p = re.sub(r'^[\\\\$\\\\{\\\\}\\\\s]*|[\\\\$\\\\{\\\\}\\\\s]*$', '', p)\n if p:\n labels.add(p)\n \nprint(\"Found labels:\")\nfor lbl in sorted(labels):\n print(f\"- {lbl}\")\n```" - } - ] - ``` - - Fix the errors and try again. - role: user - model: gpt-oss - reasoning_effort: low - stream: false - tool_choice: auto - tools: - - function: - description: |- - Execute Python code in a Docker-sandboxed environment. - - The code has access to haiku.rag functions (search, list_documents, - get_document, get_docling_document, llm) and any Python standard - library module. - - Use print() to output results. - - Structured result with success status, stdout, and stderr. - - name: execute_code - parameters: - additionalProperties: false - properties: - code: - description: Python code to execute. - type: string - required: - - code - type: object - strict: true - type: function - - function: - description: Result from RLM agent execution. - name: final_result - parameters: - additionalProperties: false - properties: - answer: - description: The answer to the user's question - type: string - program: - description: The final consolidated program - type: string - required: - - answer - - program - type: object - strict: true - type: function - uri: http://localhost:11434/v1/chat/completions - response: - headers: - content-length: - - '1389' - content-type: - - application/json - parsed_body: - choices: - - finish_reason: stop - index: 0 - message: - content: '{"answer":"The DocLayNet dataset defines the following 11 document element types (class labels):\n- Caption\n- - Footnote\n- Formula\n- List-item\n- Page-footer\n- Page-header\n- Picture\n- Section-header\n- Table\n- Text\n- - Title\n\nThese 11 are the distinct element types mentioned.\n","program":"import re\n\n# Search for the table - that lists class labels\nresults = search(''Table 1: DocLayNet dataset overview'', limit=5)\n\nlabels = set()\nfor - r in results:\n # Look for the part that lists class labels in the content\n m = re.search(r''lead us to - the definition of ([^\\.]*)\\. These ([^\\.]*)'', r[''content''])\n if m:\n # Split by comma and strip - whitespace and symbols\n parts = re.split(r'',\\s*'', m.group(2))\n for p in parts:\n p - = re.sub(r''^[\\$\\{\\}\\s]*|[\\$\\{\\}\\s]*$'', '''', p)\n if p:\n labels.add(p)\n\nprint(\"Found - labels:\")\nfor lbl in sorted(labels):\n print(f\"- {lbl}\")"}' - reasoning: Need to output JSON with answer and program fields. - role: assistant - created: 1770373456 - id: chatcmpl-841 - model: gpt-oss - object: chat.completion - system_fingerprint: fp_ollama - usage: - completion_tokens: 299 - prompt_tokens: 6665 - total_tokens: 6964 + completion_tokens: 398 + prompt_tokens: 4547 + total_tokens: 4945 status: code: 200 message: OK diff --git a/tests/cassettes/test_rlm/TestClientRLMIntegration.test_rlm_search_and_get_chunk.yaml b/tests/cassettes/test_rlm/TestClientRLMIntegration.test_rlm_search_and_get_chunk.yaml new file mode 100644 index 00000000..3e87ecb1 --- /dev/null +++ b/tests/cassettes/test_rlm/TestClientRLMIntegration.test_rlm_search_and_get_chunk.yaml @@ -0,0 +1,545 @@ +interactions: +- request: + headers: + accept: + - application/json + accept-encoding: + - gzip, deflate, zstd + connection: + - keep-alive + content-length: + - '114' + content-type: + - application/json + host: + - localhost:11434 + method: POST + parsed_body: + encoding_format: base64 + input: + - The quick brown fox jumps over the lazy dog. + model: qwen3-embedding:4b + uri: http://localhost:11434/v1/embeddings + response: + headers: + content-type: + - application/json + transfer-encoding: + - chunked + parsed_body: + data: + - embedding: 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 + index: 0 + object: embedding + model: qwen3-embedding:4b + object: list + usage: + prompt_tokens: 11 + total_tokens: 11 + status: + code: 200 + message: OK +- request: + headers: + accept: + - application/json + accept-encoding: + - gzip, deflate, zstd + connection: + - keep-alive + content-length: + - '7359' + content-type: + - application/json + host: + - localhost:11434 + method: POST + parsed_body: + messages: + - content: |- + You are a Recursive Language Model (RLM) agent that solves complex research questions by writing and executing Python code. + + IMPORTANT: You MUST use the `execute_code` tool to run Python code. The functions described below are ONLY available inside the execute_code tool - you cannot access them any other way. Always execute code to answer questions; do not just describe what code would do. + + CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just call them with `await`: + - results = await search("query") ✓ CORRECT + - from haiku.rag import search ✗ WRONG - will fail + - results = search("query") ✗ WRONG - must use await + + You have access to a sandboxed Python interpreter with these haiku.rag functions (use them directly with `await`, no imports needed): + + ## Available Functions + + ### await search(query, limit=10) -> list[dict] + Search the knowledge base using hybrid search (vector + full-text). + Returns list of dicts with keys: chunk_id, content, document_id, document_title, document_uri, score, page_numbers, headings + + ### await list_documents(limit=10, offset=0) -> list[dict] + List available documents in the knowledge base. + Returns list of dicts with keys: id, title, uri, created_at + + ### await get_document(id_or_title) -> str | None + Get the full text content of a document by ID, title, or URI. + Returns the document content as a string, or None if not found. + + ### await get_chunk(chunk_id) -> dict | None + Get a specific chunk by its ID (from search results). + Returns dict with keys: chunk_id, content, document_id, document_title, headings, page_numbers, labels + Use this to retrieve full chunk details and metadata for citation. + + ### await llm(prompt) -> str + Call an LLM directly with the given prompt. Returns the response as a string. + Use this for classification, summarization, extraction, or any task where you + already have the content and just need LLM reasoning. + + ## Pre-loaded Documents Variable + + If documents were pre-loaded for this session, a `documents` variable is available: + ```python + # documents is a list of dicts with keys: id, title, uri, content + for doc in documents: + print(doc['title'], len(doc['content'])) + ``` + Check if it exists with: `if 'documents' in dir(): ...` + + ## Available Python Features + + The interpreter supports: variables, arithmetic, strings, f-strings, lists, dicts, tuples, sets, loops, conditionals, comprehensions, functions, async/await, `map()`, `sorted()`/`.sort(key=...)`, try/except, and the `json` module. + + Not supported: imports (other than `json`), class definitions, generators/yield, match statements, decorators, `with` statements. + + For pattern matching or text extraction, use string methods (`str.split`, `str.find`, `str.startswith`, `in` operator) or the `llm()` function. + + ## Strategy Guide + + 1. **Explore First**: Start by listing documents or searching to understand what's available. Document names may differ from filenames (e.g., "tbmed593.pdf" might be stored as "TB MED 593" or similar). + 2. **If get_document returns None**: Use `await list_documents()` to see actual document titles, or `await search()` to find relevant content. + 3. **Iterative Refinement**: Run code, examine results, adjust your approach based on what you find. + 4. **Use print() Liberally**: The sandbox captures stdout - print intermediate results to see what you're working with. + 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and data structures. + 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use `await llm()`. + 7. **Cite Your Sources**: Use get_chunk() to retrieve chunk metadata for citations. Track which documents/chunks informed your answer. + + ## Example Patterns + + ### Counting documents matching a condition + ```python + docs = await list_documents(limit=100) + count = 0 + for doc in docs: + content = await get_document(doc['id']) + if content and 'keyword' in content.lower(): + count += 1 + print(f"Found in: {doc['title']}") + print(f"Total: {count}") + ``` + + ### Extracting data with llm() + ```python + numbers = [] + results = await search("financial data", limit=20) + for r in results: + extracted = await llm(f"Extract all dollar amounts from this text as a comma-separated list of numbers (no $ signs): {r['content']}") + for part in extracted.split(','): + part = part.strip().replace(',', '') + if part.isdigit(): + numbers.append(int(part)) + if numbers: + print(f"Average: {sum(numbers) / len(numbers)}") + ``` + + ### Using search results with get_chunk for citations + ```python + results = await search("safety requirements", limit=5) + for r in results: + chunk = await get_chunk(r['chunk_id']) + print(f"From '{chunk['document_title']}', page {chunk['page_numbers']}: {chunk['content'][:100]}") + ``` + + ### Using llm() for classification + ```python + content = await get_document("Q1 Report") + sentiment = await llm(f"Classify the sentiment as positive, negative, or mixed: {content}") + print(sentiment) + ``` + + ## Workflow + + 1. **ALWAYS start by using execute_code** to explore the knowledge base + 2. Run multiple code blocks as needed to gather information + 3. After collecting data, provide your final answer + + ## Output Format + + CRITICAL: Your final response MUST be valid JSON matching this exact schema: + ```json + {"answer": "Your complete answer here as a string", "program": "Your final consolidated program here as a string"} + ``` + + - `answer`: A clear answer to the user's question with key findings and references to specific documents/chunks. + - `program`: A single, self-contained Python program that produces the answer. Consolidate your exploratory code executions into one clean script. + + Do NOT return arbitrary JSON structures. Always use the exact format: {"answer": "...", "program": "..."} + + CRITICAL: You MUST call execute_code at least once before providing your answer. Never give up without trying to execute code first. + role: system + - content: Search for content about animals and tell me which document it came from. + role: user + model: gpt-oss + reasoning_effort: low + stream: false + tool_choice: auto + tools: + - function: + description: |- + Execute Python code in a sandboxed interpreter. + + The code has access to haiku.rag functions (search, list_documents, + get_document, get_chunk, llm). + + Use print() to output results. + + Structured result with success status, stdout, and stderr. + + name: execute_code + parameters: + additionalProperties: false + properties: + code: + description: Python code to execute. + type: string + required: + - code + type: object + strict: true + type: function + - function: + description: Result from RLM agent execution. + name: final_result + parameters: + additionalProperties: false + properties: + answer: + description: The answer to the user's question + type: string + program: + description: The final consolidated program + type: string + required: + - answer + - program + type: object + strict: true + type: function + uri: http://localhost:11434/v1/chat/completions + response: + headers: + content-length: + - '537' + content-type: + - application/json + parsed_body: + choices: + - finish_reason: tool_calls + index: 0 + message: + content: '' + reasoning: We need to search for animals. + role: assistant + tool_calls: + - function: + arguments: '{"code":"results = await search(\"animals\", limit=10)\nprint(results)\n"}' + name: execute_code + id: call_pvjujvr9 + index: 0 + type: function + created: 1771924521 + id: chatcmpl-217 + model: gpt-oss + object: chat.completion + system_fingerprint: fp_ollama + usage: + completion_tokens: 46 + prompt_tokens: 1629 + total_tokens: 1675 + status: + code: 200 + message: OK +- request: + headers: + accept: + - application/json + accept-encoding: + - gzip, deflate, zstd + connection: + - keep-alive + content-length: + - '77' + content-type: + - application/json + host: + - localhost:11434 + method: POST + parsed_body: + encoding_format: base64 + input: + - animals + model: qwen3-embedding:4b + uri: http://localhost:11434/v1/embeddings + response: + headers: + content-type: + - application/json + transfer-encoding: + - chunked + parsed_body: + data: + - embedding: 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 + index: 0 + object: embedding + model: qwen3-embedding:4b + object: list + usage: + prompt_tokens: 2 + total_tokens: 2 + status: + code: 200 + message: OK +- request: + headers: + accept: + - application/json + accept-encoding: + - gzip, deflate, zstd + connection: + - keep-alive + content-length: + - '8119' + content-type: + - application/json + host: + - localhost:11434 + method: POST + parsed_body: + messages: + - content: |- + You are a Recursive Language Model (RLM) agent that solves complex research questions by writing and executing Python code. + + IMPORTANT: You MUST use the `execute_code` tool to run Python code. The functions described below are ONLY available inside the execute_code tool - you cannot access them any other way. Always execute code to answer questions; do not just describe what code would do. + + CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just call them with `await`: + - results = await search("query") ✓ CORRECT + - from haiku.rag import search ✗ WRONG - will fail + - results = search("query") ✗ WRONG - must use await + + You have access to a sandboxed Python interpreter with these haiku.rag functions (use them directly with `await`, no imports needed): + + ## Available Functions + + ### await search(query, limit=10) -> list[dict] + Search the knowledge base using hybrid search (vector + full-text). + Returns list of dicts with keys: chunk_id, content, document_id, document_title, document_uri, score, page_numbers, headings + + ### await list_documents(limit=10, offset=0) -> list[dict] + List available documents in the knowledge base. + Returns list of dicts with keys: id, title, uri, created_at + + ### await get_document(id_or_title) -> str | None + Get the full text content of a document by ID, title, or URI. + Returns the document content as a string, or None if not found. + + ### await get_chunk(chunk_id) -> dict | None + Get a specific chunk by its ID (from search results). + Returns dict with keys: chunk_id, content, document_id, document_title, headings, page_numbers, labels + Use this to retrieve full chunk details and metadata for citation. + + ### await llm(prompt) -> str + Call an LLM directly with the given prompt. Returns the response as a string. + Use this for classification, summarization, extraction, or any task where you + already have the content and just need LLM reasoning. + + ## Pre-loaded Documents Variable + + If documents were pre-loaded for this session, a `documents` variable is available: + ```python + # documents is a list of dicts with keys: id, title, uri, content + for doc in documents: + print(doc['title'], len(doc['content'])) + ``` + Check if it exists with: `if 'documents' in dir(): ...` + + ## Available Python Features + + The interpreter supports: variables, arithmetic, strings, f-strings, lists, dicts, tuples, sets, loops, conditionals, comprehensions, functions, async/await, `map()`, `sorted()`/`.sort(key=...)`, try/except, and the `json` module. + + Not supported: imports (other than `json`), class definitions, generators/yield, match statements, decorators, `with` statements. + + For pattern matching or text extraction, use string methods (`str.split`, `str.find`, `str.startswith`, `in` operator) or the `llm()` function. + + ## Strategy Guide + + 1. **Explore First**: Start by listing documents or searching to understand what's available. Document names may differ from filenames (e.g., "tbmed593.pdf" might be stored as "TB MED 593" or similar). + 2. **If get_document returns None**: Use `await list_documents()` to see actual document titles, or `await search()` to find relevant content. + 3. **Iterative Refinement**: Run code, examine results, adjust your approach based on what you find. + 4. **Use print() Liberally**: The sandbox captures stdout - print intermediate results to see what you're working with. + 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and data structures. + 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use `await llm()`. + 7. **Cite Your Sources**: Use get_chunk() to retrieve chunk metadata for citations. Track which documents/chunks informed your answer. + + ## Example Patterns + + ### Counting documents matching a condition + ```python + docs = await list_documents(limit=100) + count = 0 + for doc in docs: + content = await get_document(doc['id']) + if content and 'keyword' in content.lower(): + count += 1 + print(f"Found in: {doc['title']}") + print(f"Total: {count}") + ``` + + ### Extracting data with llm() + ```python + numbers = [] + results = await search("financial data", limit=20) + for r in results: + extracted = await llm(f"Extract all dollar amounts from this text as a comma-separated list of numbers (no $ signs): {r['content']}") + for part in extracted.split(','): + part = part.strip().replace(',', '') + if part.isdigit(): + numbers.append(int(part)) + if numbers: + print(f"Average: {sum(numbers) / len(numbers)}") + ``` + + ### Using search results with get_chunk for citations + ```python + results = await search("safety requirements", limit=5) + for r in results: + chunk = await get_chunk(r['chunk_id']) + print(f"From '{chunk['document_title']}', page {chunk['page_numbers']}: {chunk['content'][:100]}") + ``` + + ### Using llm() for classification + ```python + content = await get_document("Q1 Report") + sentiment = await llm(f"Classify the sentiment as positive, negative, or mixed: {content}") + print(sentiment) + ``` + + ## Workflow + + 1. **ALWAYS start by using execute_code** to explore the knowledge base + 2. Run multiple code blocks as needed to gather information + 3. After collecting data, provide your final answer + + ## Output Format + + CRITICAL: Your final response MUST be valid JSON matching this exact schema: + ```json + {"answer": "Your complete answer here as a string", "program": "Your final consolidated program here as a string"} + ``` + + - `answer`: A clear answer to the user's question with key findings and references to specific documents/chunks. + - `program`: A single, self-contained Python program that produces the answer. Consolidate your exploratory code executions into one clean script. + + Do NOT return arbitrary JSON structures. Always use the exact format: {"answer": "...", "program": "..."} + + CRITICAL: You MUST call execute_code at least once before providing your answer. Never give up without trying to execute code first. + role: system + - content: Search for content about animals and tell me which document it came from. + role: user + - content: null + reasoning: We need to search for animals. + role: assistant + tool_calls: + - function: + arguments: '{"code":"results = await search(\"animals\", limit=10)\nprint(results)\n"}' + name: execute_code + id: call_pvjujvr9 + type: function + - content: '{"code":"results = await search(\"animals\", limit=10)\nprint(results)\n","stdout":"[{''chunk_id'': ''503004ae-51ea-4953-93c3-48acac4a929c'', + ''content'': ''The quick brown fox jumps over the lazy dog.'', ''document_id'': ''ee57cfe6-fe82-4162-afd9-f206002ae08e'', + ''document_title'': ''Animal Facts'', ''document_uri'': None, ''score'': 0.016393441706895828, ''page_numbers'': + [], ''headings'': None}]\n","stderr":"","success":true}' + role: tool + tool_call_id: call_pvjujvr9 + model: gpt-oss + reasoning_effort: low + stream: false + tool_choice: auto + tools: + - function: + description: |- + Execute Python code in a sandboxed interpreter. + + The code has access to haiku.rag functions (search, list_documents, + get_document, get_chunk, llm). + + Use print() to output results. + + Structured result with success status, stdout, and stderr. + + name: execute_code + parameters: + additionalProperties: false + properties: + code: + description: Python code to execute. + type: string + required: + - code + type: object + strict: true + type: function + - function: + description: Result from RLM agent execution. + name: final_result + parameters: + additionalProperties: false + properties: + answer: + description: The answer to the user's question + type: string + program: + description: The final consolidated program + type: string + required: + - answer + - program + type: object + strict: true + type: function + uri: http://localhost:11434/v1/chat/completions + response: + headers: + content-length: + - '888' + content-type: + - application/json + parsed_body: + choices: + - finish_reason: stop + index: 0 + message: + content: '{"answer":"The document containing content about animals is \"Animal Facts\" (document ID: ee57cfe6-fe82-4162-afd9-f206002ae08e).","program":"import + asyncio\n\nasync def main():\n results = await search(\"animals\", limit=10)\n if results:\n chunk + = results[0]\n print(f\"Document: {chunk[''document_title'']} (ID: {chunk[''document_id'']})\")\n else:\n print(\"No + animal-related content found.\")\n\nasyncio.run(main())"}' + reasoning: It found one chunk. We need to give answer with source. Provide program that searches and prints answer. + role: assistant + created: 1771924525 + id: chatcmpl-109 + model: gpt-oss + object: chat.completion + system_fingerprint: fp_ollama + usage: + completion_tokens: 168 + prompt_tokens: 1822 + total_tokens: 1990 + status: + code: 200 + message: OK +version: 1 diff --git a/tests/cassettes/test_rlm/TestClientRLMIntegration.test_rlm_semantic_analysis_with_llm.yaml b/tests/cassettes/test_rlm/TestClientRLMIntegration.test_rlm_semantic_analysis_with_llm.yaml index 0e2b9c9d..375c1602 100644 --- a/tests/cassettes/test_rlm/TestClientRLMIntegration.test_rlm_semantic_analysis_with_llm.yaml +++ b/tests/cassettes/test_rlm/TestClientRLMIntegration.test_rlm_semantic_analysis_with_llm.yaml @@ -131,7 +131,7 @@ interactions: connection: - keep-alive content-length: - - '7836' + - '7387' content-type: - application/json host: @@ -144,32 +144,33 @@ interactions: IMPORTANT: You MUST use the `execute_code` tool to run Python code. The functions described below are ONLY available inside the execute_code tool - you cannot access them any other way. Always execute code to answer questions; do not just describe what code would do. - CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just use them directly: - - search("query") ✓ CORRECT + CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just call them with `await`: + - results = await search("query") ✓ CORRECT - from haiku.rag import search ✗ WRONG - will fail + - results = search("query") ✗ WRONG - must use await - You have access to a sandboxed Python environment with these haiku.rag functions (use them directly, no imports needed): + You have access to a sandboxed Python interpreter with these haiku.rag functions (use them directly with `await`, no imports needed): ## Available Functions - ### search(query, limit=10) -> list[dict] + ### await search(query, limit=10) -> list[dict] Search the knowledge base using hybrid search (vector + full-text). Returns list of dicts with keys: chunk_id, content, document_id, document_title, document_uri, score, page_numbers, headings - ### list_documents(limit=10, offset=0) -> list[dict] + ### await list_documents(limit=10, offset=0) -> list[dict] List available documents in the knowledge base. Returns list of dicts with keys: id, title, uri, created_at - ### get_document(id_or_title) -> str | None + ### await get_document(id_or_title) -> str | None Get the full text content of a document by ID, title, or URI. Returns the document content as a string, or None if not found. - ### get_docling_document(id_or_title) -> DoclingDocument | None - Get the structured DoclingDocument object for advanced analysis. - Returns a DoclingDocument object, or None if not found. - See "DoclingDocument API" section below for how to use it. + ### await get_chunk(chunk_id) -> dict | None + Get a specific chunk by its ID (from search results). + Returns dict with keys: chunk_id, content, document_id, document_title, headings, page_numbers, labels + Use this to retrieve full chunk details and metadata for citation. - ### llm(prompt) -> str + ### await llm(prompt) -> str Call an LLM directly with the given prompt. Returns the response as a string. Use this for classification, summarization, extraction, or any task where you already have the content and just need LLM reasoning. @@ -184,93 +185,64 @@ interactions: ``` Check if it exists with: `if 'documents' in dir(): ...` - ## Standard Library Modules - You can import any Python standard library module. + ## Available Python Features + + The interpreter supports: variables, arithmetic, strings, f-strings, lists, dicts, tuples, sets, loops, conditionals, comprehensions, functions, async/await, `map()`, `sorted()`/`.sort(key=...)`, try/except, and the `json` module. + + Not supported: imports (other than `json`), class definitions, generators/yield, match statements, decorators, `with` statements. + + For pattern matching or text extraction, use string methods (`str.split`, `str.find`, `str.startswith`, `in` operator) or the `llm()` function. ## Strategy Guide 1. **Explore First**: Start by listing documents or searching to understand what's available. Document names may differ from filenames (e.g., "tbmed593.pdf" might be stored as "TB MED 593" or similar). - 2. **If get_document returns None**: Use `list_documents()` to see actual document titles, or `search()` to find relevant content. + 2. **If get_document returns None**: Use `await list_documents()` to see actual document titles, or `await search()` to find relevant content. 3. **Iterative Refinement**: Run code, examine results, adjust your approach based on what you find. - 4. **Use print() Liberally**: The REPL captures stdout - print intermediate results to see what you're working with. - 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and use collections. - 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use llm(). - 7. **Cite Your Sources**: Track which documents/chunks informed your answer for citation. - - ## DoclingDocument API - - When you call `get_docling_document(id_or_title)`, you get a DoclingDocument object for structured document analysis. - - ### Properties - - `doc.texts` - List of all text items (paragraphs, headings, etc.) - - `doc.tables` - List of all tables - - `doc.pictures` - List of all pictures/figures - - `doc.name` - Document name - - ### Methods - - `doc.iterate_items(with_groups=False)` - Iterate all items with hierarchy level - Returns tuples of (item, level) where level is nesting depth - - `doc.export_to_markdown()` - Export entire document as markdown string - - ### Text Item Properties - - `item.text` - The text content - - `item.label` - Type: title, paragraph, section_header, list_item, etc. (lowercase enum values) - - `item.prov` - Provenance (page numbers, bounding boxes) - - ### Table Access - - `table.data.num_rows`, `table.data.num_cols` - Dimensions - - `table.data.table_cells` - List of TableCell objects - - `cell.text`, `cell.start_row_offset_idx`, `cell.start_col_offset_idx` - - ### Example Usage - ```python - doc = get_docling_document("My Document") - - # Get all headings - headings = [t.text for t in doc.texts if "header" in str(t.label)] - - # Iterate with structure - for item, level in doc.iterate_items(): - print(" " * level + item.text[:50]) - - # Extract table data - for table in doc.tables: - for cell in table.data.table_cells: - print(f"Row {cell.start_row_offset_idx}, Col {cell.start_col_offset_idx}: {cell.text}") - ``` + 4. **Use print() Liberally**: The sandbox captures stdout - print intermediate results to see what you're working with. + 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and data structures. + 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use `await llm()`. + 7. **Cite Your Sources**: Use get_chunk() to retrieve chunk metadata for citations. Track which documents/chunks informed your answer. ## Example Patterns ### Counting documents matching a condition ```python - docs = list_documents(limit=100) + docs = await list_documents(limit=100) count = 0 for doc in docs: - content = get_document(doc['id']) + content = await get_document(doc['id']) if content and 'keyword' in content.lower(): count += 1 print(f"Found in: {doc['title']}") print(f"Total: {count}") ``` - ### Aggregating data across documents + ### Extracting data with llm() ```python - import re numbers = [] - results = search("financial data", limit=20) + results = await search("financial data", limit=20) for r in results: - matches = re.findall(r'\$([\d,]+)', r['content']) - for m in matches: - numbers.append(int(m.replace(',', ''))) - print(f"Average: ${sum(numbers)/len(numbers):,.2f}") + extracted = await llm(f"Extract all dollar amounts from this text as a comma-separated list of numbers (no $ signs): {r['content']}") + for part in extracted.split(','): + part = part.strip().replace(',', '') + if part.isdigit(): + numbers.append(int(part)) + if numbers: + print(f"Average: {sum(numbers) / len(numbers)}") + ``` + + ### Using search results with get_chunk for citations + ```python + results = await search("safety requirements", limit=5) + for r in results: + chunk = await get_chunk(r['chunk_id']) + print(f"From '{chunk['document_title']}', page {chunk['page_numbers']}: {chunk['content'][:100]}") ``` ### Using llm() for classification ```python - # Get document content - content = get_document("Q1 Report") - # Use llm() to classify sentiment - sentiment = llm(f"Classify the sentiment as positive, negative, or mixed: {content}") + content = await get_document("Q1 Report") + sentiment = await llm(f"Classify the sentiment as positive, negative, or mixed: {content}") print(sentiment) ``` @@ -303,11 +275,10 @@ interactions: tools: - function: description: |- - Execute Python code in a Docker-sandboxed environment. + Execute Python code in a sandboxed interpreter. The code has access to haiku.rag functions (search, list_documents, - get_document, get_docling_document, llm) and any Python standard - library module. + get_document, get_chunk, llm). Use print() to output results. @@ -347,7 +318,7 @@ interactions: response: headers: content-length: - - '592' + - '677' content-type: - application/json parsed_body: @@ -356,24 +327,25 @@ interactions: index: 0 message: content: '' - reasoning: Need to search for quarterly updates documents. Use search. + reasoning: Need to search for quarterly updates. Let's search "quarterly update". role: assistant tool_calls: - function: - arguments: '{"code":"results = search(\"quarterly update\", limit=50)\nprint(len(results))\nprint(results[:3])"}' + arguments: '{"code":"results = await search(\"quarterly update\", limit=20)\nprint(len(results))\nfor r in results[:5]:\n print(r[''document_title''], + r[''chunk_id''], r[''score''])\n"}' name: execute_code - id: call_vnbck7g3 + id: call_dwvve4zt index: 0 type: function - created: 1770373405 - id: chatcmpl-477 + created: 1771924530 + id: chatcmpl-596 model: gpt-oss object: chat.completion system_fingerprint: fp_ollama usage: - completion_tokens: 58 - prompt_tokens: 1758 - total_tokens: 1816 + completion_tokens: 84 + prompt_tokens: 1634 + total_tokens: 1718 status: code: 200 message: OK @@ -386,7 +358,47 @@ interactions: connection: - keep-alive content-length: - - '9608' + - '86' + content-type: + - application/json + host: + - localhost:11434 + method: POST + parsed_body: + encoding_format: base64 + input: + - quarterly update + model: qwen3-embedding:4b + uri: http://localhost:11434/v1/embeddings + response: + headers: + content-type: + - application/json + transfer-encoding: + - chunked + parsed_body: + data: + - embedding: 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 + index: 0 + object: embedding + model: qwen3-embedding:4b + object: list + usage: + prompt_tokens: 4 + total_tokens: 4 + status: + code: 200 + message: OK +- request: + headers: + accept: + - application/json + accept-encoding: + - gzip, deflate, zstd + connection: + - keep-alive + content-length: + - '8302' content-type: - application/json host: @@ -399,32 +411,33 @@ interactions: IMPORTANT: You MUST use the `execute_code` tool to run Python code. The functions described below are ONLY available inside the execute_code tool - you cannot access them any other way. Always execute code to answer questions; do not just describe what code would do. - CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just use them directly: - - search("query") ✓ CORRECT + CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just call them with `await`: + - results = await search("query") ✓ CORRECT - from haiku.rag import search ✗ WRONG - will fail + - results = search("query") ✗ WRONG - must use await - You have access to a sandboxed Python environment with these haiku.rag functions (use them directly, no imports needed): + You have access to a sandboxed Python interpreter with these haiku.rag functions (use them directly with `await`, no imports needed): ## Available Functions - ### search(query, limit=10) -> list[dict] + ### await search(query, limit=10) -> list[dict] Search the knowledge base using hybrid search (vector + full-text). Returns list of dicts with keys: chunk_id, content, document_id, document_title, document_uri, score, page_numbers, headings - ### list_documents(limit=10, offset=0) -> list[dict] + ### await list_documents(limit=10, offset=0) -> list[dict] List available documents in the knowledge base. Returns list of dicts with keys: id, title, uri, created_at - ### get_document(id_or_title) -> str | None + ### await get_document(id_or_title) -> str | None Get the full text content of a document by ID, title, or URI. Returns the document content as a string, or None if not found. - ### get_docling_document(id_or_title) -> DoclingDocument | None - Get the structured DoclingDocument object for advanced analysis. - Returns a DoclingDocument object, or None if not found. - See "DoclingDocument API" section below for how to use it. + ### await get_chunk(chunk_id) -> dict | None + Get a specific chunk by its ID (from search results). + Returns dict with keys: chunk_id, content, document_id, document_title, headings, page_numbers, labels + Use this to retrieve full chunk details and metadata for citation. - ### llm(prompt) -> str + ### await llm(prompt) -> str Call an LLM directly with the given prompt. Returns the response as a string. Use this for classification, summarization, extraction, or any task where you already have the content and just need LLM reasoning. @@ -439,93 +452,64 @@ interactions: ``` Check if it exists with: `if 'documents' in dir(): ...` - ## Standard Library Modules - You can import any Python standard library module. + ## Available Python Features + + The interpreter supports: variables, arithmetic, strings, f-strings, lists, dicts, tuples, sets, loops, conditionals, comprehensions, functions, async/await, `map()`, `sorted()`/`.sort(key=...)`, try/except, and the `json` module. + + Not supported: imports (other than `json`), class definitions, generators/yield, match statements, decorators, `with` statements. + + For pattern matching or text extraction, use string methods (`str.split`, `str.find`, `str.startswith`, `in` operator) or the `llm()` function. ## Strategy Guide 1. **Explore First**: Start by listing documents or searching to understand what's available. Document names may differ from filenames (e.g., "tbmed593.pdf" might be stored as "TB MED 593" or similar). - 2. **If get_document returns None**: Use `list_documents()` to see actual document titles, or `search()` to find relevant content. + 2. **If get_document returns None**: Use `await list_documents()` to see actual document titles, or `await search()` to find relevant content. 3. **Iterative Refinement**: Run code, examine results, adjust your approach based on what you find. - 4. **Use print() Liberally**: The REPL captures stdout - print intermediate results to see what you're working with. - 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and use collections. - 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use llm(). - 7. **Cite Your Sources**: Track which documents/chunks informed your answer for citation. - - ## DoclingDocument API - - When you call `get_docling_document(id_or_title)`, you get a DoclingDocument object for structured document analysis. - - ### Properties - - `doc.texts` - List of all text items (paragraphs, headings, etc.) - - `doc.tables` - List of all tables - - `doc.pictures` - List of all pictures/figures - - `doc.name` - Document name - - ### Methods - - `doc.iterate_items(with_groups=False)` - Iterate all items with hierarchy level - Returns tuples of (item, level) where level is nesting depth - - `doc.export_to_markdown()` - Export entire document as markdown string - - ### Text Item Properties - - `item.text` - The text content - - `item.label` - Type: title, paragraph, section_header, list_item, etc. (lowercase enum values) - - `item.prov` - Provenance (page numbers, bounding boxes) - - ### Table Access - - `table.data.num_rows`, `table.data.num_cols` - Dimensions - - `table.data.table_cells` - List of TableCell objects - - `cell.text`, `cell.start_row_offset_idx`, `cell.start_col_offset_idx` - - ### Example Usage - ```python - doc = get_docling_document("My Document") - - # Get all headings - headings = [t.text for t in doc.texts if "header" in str(t.label)] - - # Iterate with structure - for item, level in doc.iterate_items(): - print(" " * level + item.text[:50]) - - # Extract table data - for table in doc.tables: - for cell in table.data.table_cells: - print(f"Row {cell.start_row_offset_idx}, Col {cell.start_col_offset_idx}: {cell.text}") - ``` + 4. **Use print() Liberally**: The sandbox captures stdout - print intermediate results to see what you're working with. + 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and data structures. + 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use `await llm()`. + 7. **Cite Your Sources**: Use get_chunk() to retrieve chunk metadata for citations. Track which documents/chunks informed your answer. ## Example Patterns ### Counting documents matching a condition ```python - docs = list_documents(limit=100) + docs = await list_documents(limit=100) count = 0 for doc in docs: - content = get_document(doc['id']) + content = await get_document(doc['id']) if content and 'keyword' in content.lower(): count += 1 print(f"Found in: {doc['title']}") print(f"Total: {count}") ``` - ### Aggregating data across documents + ### Extracting data with llm() ```python - import re numbers = [] - results = search("financial data", limit=20) + results = await search("financial data", limit=20) for r in results: - matches = re.findall(r'\$([\d,]+)', r['content']) - for m in matches: - numbers.append(int(m.replace(',', ''))) - print(f"Average: ${sum(numbers)/len(numbers):,.2f}") + extracted = await llm(f"Extract all dollar amounts from this text as a comma-separated list of numbers (no $ signs): {r['content']}") + for part in extracted.split(','): + part = part.strip().replace(',', '') + if part.isdigit(): + numbers.append(int(part)) + if numbers: + print(f"Average: {sum(numbers) / len(numbers)}") + ``` + + ### Using search results with get_chunk for citations + ```python + results = await search("safety requirements", limit=5) + for r in results: + chunk = await get_chunk(r['chunk_id']) + print(f"From '{chunk['document_title']}', page {chunk['page_numbers']}: {chunk['content'][:100]}") ``` ### Using llm() for classification ```python - # Get document content - content = get_document("Q1 Report") - # Use llm() to classify sentiment - sentiment = llm(f"Classify the sentiment as positive, negative, or mixed: {content}") + content = await get_document("Q1 Report") + sentiment = await llm(f"Classify the sentiment as positive, negative, or mixed: {content}") print(sentiment) ``` @@ -551,31 +535,22 @@ interactions: role: system - content: Analyze the sentiment of each quarterly update. How many quarters were positive, negative, and mixed? role: user - - content: |- - - Need to search for quarterly updates documents. Use search. - + - content: null + reasoning: Need to search for quarterly updates. Let's search "quarterly update". role: assistant tool_calls: - function: - arguments: '{"code":"results = search(\"quarterly update\", limit=50)\nprint(len(results))\nprint(results[:3])"}' + arguments: '{"code":"results = await search(\"quarterly update\", limit=20)\nprint(len(results))\nfor r in results[:5]:\n print(r[''document_title''], + r[''chunk_id''], r[''score''])\n"}' name: execute_code - id: call_vnbck7g3 + id: call_dwvve4zt type: function - - content: '{"code":"results = search(\"quarterly update\", limit=50)\nprint(len(results))\nprint(results[:3])","stdout":"3\n[{''chunk_id'': - ''f43599fd-f1ef-456d-8aee-28467c9e467b'', ''content'': ''Mixed results this quarter. While product quality improved, - marketing campaigns underperformed. Revenue was flat compared to last year but customer retention increased.'', - ''document_id'': ''1c58eccc-aad4-4b4e-8632-854dfbafc78e'', ''document_title'': ''Q3 Update'', ''document_uri'': - None, ''score'': 0.032786883413791656, ''page_numbers'': [], ''headings'': None}, {''chunk_id'': ''ce740ef8-ae7f-4efb-81c1-1ffa34323f56'', - ''content'': ''We faced significant challenges this quarter. Supply chain issues caused delays, and we missed our - revenue target by 15%. Several key employees left the company.'', ''document_id'': ''afb82f72-2eb0-4d3e-ad45-95c641e8cb71'', - ''document_title'': ''Q2 Update'', ''document_uri'': None, ''score'': 0.032258063554763794, ''page_numbers'': [], - ''headings'': None}, {''chunk_id'': ''77a36420-7c61-4881-bd05-ba7aec532934'', ''content'': ''The new product launch - exceeded expectations. Sales grew 40% and customer feedback has been overwhelmingly positive. Team morale is at - an all-time high.'', ''document_id'': ''bea81446-5d06-489f-9309-e97aa9762a96'', ''document_title'': ''Q1 Update'', - ''document_uri'': None, ''score'': 0.01587301678955555, ''page_numbers'': [], ''headings'': None}]\n","stderr":"","success":true}' + - content: '{"code":"results = await search(\"quarterly update\", limit=20)\nprint(len(results))\nfor r in results[:5]:\n print(r[''document_title''], + r[''chunk_id''], r[''score''])\n","stdout":"3\nQ3 Update d651d341-7708-4ff3-9d4b-54bcd41694bf 0.032786883413791656\nQ2 + Update f7a3ff56-4f07-479b-8320-62a3b6a2bf48 0.032258063554763794\nQ1 Update 6797ade9-2418-4487-a3ee-bcebb62b9cf9 + 0.01587301678955555\n","stderr":"","success":true}' role: tool - tool_call_id: call_vnbck7g3 + tool_call_id: call_dwvve4zt model: gpt-oss reasoning_effort: low stream: false @@ -583,11 +558,10 @@ interactions: tools: - function: description: |- - Execute Python code in a Docker-sandboxed environment. + Execute Python code in a sandboxed interpreter. The code has access to haiku.rag functions (search, list_documents, - get_document, get_docling_document, llm) and any Python standard - library module. + get_document, get_chunk, llm). Use print() to output results. @@ -627,7 +601,968 @@ interactions: response: headers: content-length: - - '1333' + - '707' + content-type: + - application/json + parsed_body: + choices: + - finish_reason: tool_calls + index: 0 + message: + content: '' + reasoning: We have Q3, Q2, Q1. Need sentiment classify each. Use llm on content. get_chunk each. Let's get chunks. + role: assistant + tool_calls: + - function: + arguments: '{"code":"chunks = []\nfor r in results:\n chunk = await get_chunk(r[''chunk_id''])\n chunks.append(chunk)\n print(chunk[''document_title''], + len(chunk[''content'']))\n"}' + name: execute_code + id: call_juxb6n6o + index: 0 + type: function + created: 1771924532 + id: chatcmpl-716 + model: gpt-oss + object: chat.completion + system_fingerprint: fp_ollama + usage: + completion_tokens: 97 + prompt_tokens: 1901 + total_tokens: 1998 + status: + code: 200 + message: OK +- request: + headers: + accept: + - application/json + accept-encoding: + - gzip, deflate, zstd + connection: + - keep-alive + content-length: + - '9074' + content-type: + - application/json + host: + - localhost:11434 + method: POST + parsed_body: + messages: + - content: |- + You are a Recursive Language Model (RLM) agent that solves complex research questions by writing and executing Python code. + + IMPORTANT: You MUST use the `execute_code` tool to run Python code. The functions described below are ONLY available inside the execute_code tool - you cannot access them any other way. Always execute code to answer questions; do not just describe what code would do. + + CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just call them with `await`: + - results = await search("query") ✓ CORRECT + - from haiku.rag import search ✗ WRONG - will fail + - results = search("query") ✗ WRONG - must use await + + You have access to a sandboxed Python interpreter with these haiku.rag functions (use them directly with `await`, no imports needed): + + ## Available Functions + + ### await search(query, limit=10) -> list[dict] + Search the knowledge base using hybrid search (vector + full-text). + Returns list of dicts with keys: chunk_id, content, document_id, document_title, document_uri, score, page_numbers, headings + + ### await list_documents(limit=10, offset=0) -> list[dict] + List available documents in the knowledge base. + Returns list of dicts with keys: id, title, uri, created_at + + ### await get_document(id_or_title) -> str | None + Get the full text content of a document by ID, title, or URI. + Returns the document content as a string, or None if not found. + + ### await get_chunk(chunk_id) -> dict | None + Get a specific chunk by its ID (from search results). + Returns dict with keys: chunk_id, content, document_id, document_title, headings, page_numbers, labels + Use this to retrieve full chunk details and metadata for citation. + + ### await llm(prompt) -> str + Call an LLM directly with the given prompt. Returns the response as a string. + Use this for classification, summarization, extraction, or any task where you + already have the content and just need LLM reasoning. + + ## Pre-loaded Documents Variable + + If documents were pre-loaded for this session, a `documents` variable is available: + ```python + # documents is a list of dicts with keys: id, title, uri, content + for doc in documents: + print(doc['title'], len(doc['content'])) + ``` + Check if it exists with: `if 'documents' in dir(): ...` + + ## Available Python Features + + The interpreter supports: variables, arithmetic, strings, f-strings, lists, dicts, tuples, sets, loops, conditionals, comprehensions, functions, async/await, `map()`, `sorted()`/`.sort(key=...)`, try/except, and the `json` module. + + Not supported: imports (other than `json`), class definitions, generators/yield, match statements, decorators, `with` statements. + + For pattern matching or text extraction, use string methods (`str.split`, `str.find`, `str.startswith`, `in` operator) or the `llm()` function. + + ## Strategy Guide + + 1. **Explore First**: Start by listing documents or searching to understand what's available. Document names may differ from filenames (e.g., "tbmed593.pdf" might be stored as "TB MED 593" or similar). + 2. **If get_document returns None**: Use `await list_documents()` to see actual document titles, or `await search()` to find relevant content. + 3. **Iterative Refinement**: Run code, examine results, adjust your approach based on what you find. + 4. **Use print() Liberally**: The sandbox captures stdout - print intermediate results to see what you're working with. + 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and data structures. + 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use `await llm()`. + 7. **Cite Your Sources**: Use get_chunk() to retrieve chunk metadata for citations. Track which documents/chunks informed your answer. + + ## Example Patterns + + ### Counting documents matching a condition + ```python + docs = await list_documents(limit=100) + count = 0 + for doc in docs: + content = await get_document(doc['id']) + if content and 'keyword' in content.lower(): + count += 1 + print(f"Found in: {doc['title']}") + print(f"Total: {count}") + ``` + + ### Extracting data with llm() + ```python + numbers = [] + results = await search("financial data", limit=20) + for r in results: + extracted = await llm(f"Extract all dollar amounts from this text as a comma-separated list of numbers (no $ signs): {r['content']}") + for part in extracted.split(','): + part = part.strip().replace(',', '') + if part.isdigit(): + numbers.append(int(part)) + if numbers: + print(f"Average: {sum(numbers) / len(numbers)}") + ``` + + ### Using search results with get_chunk for citations + ```python + results = await search("safety requirements", limit=5) + for r in results: + chunk = await get_chunk(r['chunk_id']) + print(f"From '{chunk['document_title']}', page {chunk['page_numbers']}: {chunk['content'][:100]}") + ``` + + ### Using llm() for classification + ```python + content = await get_document("Q1 Report") + sentiment = await llm(f"Classify the sentiment as positive, negative, or mixed: {content}") + print(sentiment) + ``` + + ## Workflow + + 1. **ALWAYS start by using execute_code** to explore the knowledge base + 2. Run multiple code blocks as needed to gather information + 3. After collecting data, provide your final answer + + ## Output Format + + CRITICAL: Your final response MUST be valid JSON matching this exact schema: + ```json + {"answer": "Your complete answer here as a string", "program": "Your final consolidated program here as a string"} + ``` + + - `answer`: A clear answer to the user's question with key findings and references to specific documents/chunks. + - `program`: A single, self-contained Python program that produces the answer. Consolidate your exploratory code executions into one clean script. + + Do NOT return arbitrary JSON structures. Always use the exact format: {"answer": "...", "program": "..."} + + CRITICAL: You MUST call execute_code at least once before providing your answer. Never give up without trying to execute code first. + role: system + - content: Analyze the sentiment of each quarterly update. How many quarters were positive, negative, and mixed? + role: user + - content: null + reasoning: Need to search for quarterly updates. Let's search "quarterly update". + role: assistant + tool_calls: + - function: + arguments: '{"code":"results = await search(\"quarterly update\", limit=20)\nprint(len(results))\nfor r in results[:5]:\n print(r[''document_title''], + r[''chunk_id''], r[''score''])\n"}' + name: execute_code + id: call_dwvve4zt + type: function + - content: '{"code":"results = await search(\"quarterly update\", limit=20)\nprint(len(results))\nfor r in results[:5]:\n print(r[''document_title''], + r[''chunk_id''], r[''score''])\n","stdout":"3\nQ3 Update d651d341-7708-4ff3-9d4b-54bcd41694bf 0.032786883413791656\nQ2 + Update f7a3ff56-4f07-479b-8320-62a3b6a2bf48 0.032258063554763794\nQ1 Update 6797ade9-2418-4487-a3ee-bcebb62b9cf9 + 0.01587301678955555\n","stderr":"","success":true}' + role: tool + tool_call_id: call_dwvve4zt + - content: null + reasoning: We have Q3, Q2, Q1. Need sentiment classify each. Use llm on content. get_chunk each. Let's get chunks. + role: assistant + tool_calls: + - function: + arguments: '{"code":"chunks = []\nfor r in results:\n chunk = await get_chunk(r[''chunk_id''])\n chunks.append(chunk)\n print(chunk[''document_title''], + len(chunk[''content'']))\n"}' + name: execute_code + id: call_juxb6n6o + type: function + - content: '{"code":"chunks = []\nfor r in results:\n chunk = await get_chunk(r[''chunk_id''])\n chunks.append(chunk)\n print(chunk[''document_title''], + len(chunk[''content'']))\n","stdout":"","stderr":"NameError: name ''results'' is not defined","success":false}' + role: tool + tool_call_id: call_juxb6n6o + model: gpt-oss + reasoning_effort: low + stream: false + tool_choice: auto + tools: + - function: + description: |- + Execute Python code in a sandboxed interpreter. + + The code has access to haiku.rag functions (search, list_documents, + get_document, get_chunk, llm). + + Use print() to output results. + + Structured result with success status, stdout, and stderr. + + name: execute_code + parameters: + additionalProperties: false + properties: + code: + description: Python code to execute. + type: string + required: + - code + type: object + strict: true + type: function + - function: + description: Result from RLM agent execution. + name: final_result + parameters: + additionalProperties: false + properties: + answer: + description: The answer to the user's question + type: string + program: + description: The final consolidated program + type: string + required: + - answer + - program + type: object + strict: true + type: function + uri: http://localhost:11434/v1/chat/completions + response: + headers: + content-length: + - '762' + content-type: + - application/json + parsed_body: + choices: + - finish_reason: tool_calls + index: 0 + message: + content: '' + reasoning: Need to keep results across calls? Variables persist? Might need to re-search. Let's re-execute. + role: assistant + tool_calls: + - function: + arguments: '{"code":"results = await search(\"quarterly update\", limit=20)\nchunks = []\nfor r in results:\n chunk + = await get_chunk(r[''chunk_id''])\n chunks.append(chunk)\n print(chunk[''document_title''], len(chunk[''content'']))\n"}' + name: execute_code + id: call_7m5eb6ho + index: 0 + type: function + created: 1771924534 + id: chatcmpl-681 + model: gpt-oss + object: chat.completion + system_fingerprint: fp_ollama + usage: + completion_tokens: 102 + prompt_tokens: 2076 + total_tokens: 2178 + status: + code: 200 + message: OK +- request: + headers: + accept: + - application/json + accept-encoding: + - gzip, deflate, zstd + connection: + - keep-alive + content-length: + - '86' + content-type: + - application/json + host: + - localhost:11434 + method: POST + parsed_body: + encoding_format: base64 + input: + - quarterly update + model: qwen3-embedding:4b + uri: http://localhost:11434/v1/embeddings + response: + headers: + content-type: + - application/json + transfer-encoding: + - chunked + parsed_body: + data: + - embedding: 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 + index: 0 + object: embedding + model: qwen3-embedding:4b + object: list + usage: + prompt_tokens: 4 + total_tokens: 4 + status: + code: 200 + message: OK +- request: + headers: + accept: + - application/json + accept-encoding: + - gzip, deflate, zstd + connection: + - keep-alive + content-length: + - '9968' + content-type: + - application/json + host: + - localhost:11434 + method: POST + parsed_body: + messages: + - content: |- + You are a Recursive Language Model (RLM) agent that solves complex research questions by writing and executing Python code. + + IMPORTANT: You MUST use the `execute_code` tool to run Python code. The functions described below are ONLY available inside the execute_code tool - you cannot access them any other way. Always execute code to answer questions; do not just describe what code would do. + + CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just call them with `await`: + - results = await search("query") ✓ CORRECT + - from haiku.rag import search ✗ WRONG - will fail + - results = search("query") ✗ WRONG - must use await + + You have access to a sandboxed Python interpreter with these haiku.rag functions (use them directly with `await`, no imports needed): + + ## Available Functions + + ### await search(query, limit=10) -> list[dict] + Search the knowledge base using hybrid search (vector + full-text). + Returns list of dicts with keys: chunk_id, content, document_id, document_title, document_uri, score, page_numbers, headings + + ### await list_documents(limit=10, offset=0) -> list[dict] + List available documents in the knowledge base. + Returns list of dicts with keys: id, title, uri, created_at + + ### await get_document(id_or_title) -> str | None + Get the full text content of a document by ID, title, or URI. + Returns the document content as a string, or None if not found. + + ### await get_chunk(chunk_id) -> dict | None + Get a specific chunk by its ID (from search results). + Returns dict with keys: chunk_id, content, document_id, document_title, headings, page_numbers, labels + Use this to retrieve full chunk details and metadata for citation. + + ### await llm(prompt) -> str + Call an LLM directly with the given prompt. Returns the response as a string. + Use this for classification, summarization, extraction, or any task where you + already have the content and just need LLM reasoning. + + ## Pre-loaded Documents Variable + + If documents were pre-loaded for this session, a `documents` variable is available: + ```python + # documents is a list of dicts with keys: id, title, uri, content + for doc in documents: + print(doc['title'], len(doc['content'])) + ``` + Check if it exists with: `if 'documents' in dir(): ...` + + ## Available Python Features + + The interpreter supports: variables, arithmetic, strings, f-strings, lists, dicts, tuples, sets, loops, conditionals, comprehensions, functions, async/await, `map()`, `sorted()`/`.sort(key=...)`, try/except, and the `json` module. + + Not supported: imports (other than `json`), class definitions, generators/yield, match statements, decorators, `with` statements. + + For pattern matching or text extraction, use string methods (`str.split`, `str.find`, `str.startswith`, `in` operator) or the `llm()` function. + + ## Strategy Guide + + 1. **Explore First**: Start by listing documents or searching to understand what's available. Document names may differ from filenames (e.g., "tbmed593.pdf" might be stored as "TB MED 593" or similar). + 2. **If get_document returns None**: Use `await list_documents()` to see actual document titles, or `await search()` to find relevant content. + 3. **Iterative Refinement**: Run code, examine results, adjust your approach based on what you find. + 4. **Use print() Liberally**: The sandbox captures stdout - print intermediate results to see what you're working with. + 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and data structures. + 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use `await llm()`. + 7. **Cite Your Sources**: Use get_chunk() to retrieve chunk metadata for citations. Track which documents/chunks informed your answer. + + ## Example Patterns + + ### Counting documents matching a condition + ```python + docs = await list_documents(limit=100) + count = 0 + for doc in docs: + content = await get_document(doc['id']) + if content and 'keyword' in content.lower(): + count += 1 + print(f"Found in: {doc['title']}") + print(f"Total: {count}") + ``` + + ### Extracting data with llm() + ```python + numbers = [] + results = await search("financial data", limit=20) + for r in results: + extracted = await llm(f"Extract all dollar amounts from this text as a comma-separated list of numbers (no $ signs): {r['content']}") + for part in extracted.split(','): + part = part.strip().replace(',', '') + if part.isdigit(): + numbers.append(int(part)) + if numbers: + print(f"Average: {sum(numbers) / len(numbers)}") + ``` + + ### Using search results with get_chunk for citations + ```python + results = await search("safety requirements", limit=5) + for r in results: + chunk = await get_chunk(r['chunk_id']) + print(f"From '{chunk['document_title']}', page {chunk['page_numbers']}: {chunk['content'][:100]}") + ``` + + ### Using llm() for classification + ```python + content = await get_document("Q1 Report") + sentiment = await llm(f"Classify the sentiment as positive, negative, or mixed: {content}") + print(sentiment) + ``` + + ## Workflow + + 1. **ALWAYS start by using execute_code** to explore the knowledge base + 2. Run multiple code blocks as needed to gather information + 3. After collecting data, provide your final answer + + ## Output Format + + CRITICAL: Your final response MUST be valid JSON matching this exact schema: + ```json + {"answer": "Your complete answer here as a string", "program": "Your final consolidated program here as a string"} + ``` + + - `answer`: A clear answer to the user's question with key findings and references to specific documents/chunks. + - `program`: A single, self-contained Python program that produces the answer. Consolidate your exploratory code executions into one clean script. + + Do NOT return arbitrary JSON structures. Always use the exact format: {"answer": "...", "program": "..."} + + CRITICAL: You MUST call execute_code at least once before providing your answer. Never give up without trying to execute code first. + role: system + - content: Analyze the sentiment of each quarterly update. How many quarters were positive, negative, and mixed? + role: user + - content: null + reasoning: Need to search for quarterly updates. Let's search "quarterly update". + role: assistant + tool_calls: + - function: + arguments: '{"code":"results = await search(\"quarterly update\", limit=20)\nprint(len(results))\nfor r in results[:5]:\n print(r[''document_title''], + r[''chunk_id''], r[''score''])\n"}' + name: execute_code + id: call_dwvve4zt + type: function + - content: '{"code":"results = await search(\"quarterly update\", limit=20)\nprint(len(results))\nfor r in results[:5]:\n print(r[''document_title''], + r[''chunk_id''], r[''score''])\n","stdout":"3\nQ3 Update d651d341-7708-4ff3-9d4b-54bcd41694bf 0.032786883413791656\nQ2 + Update f7a3ff56-4f07-479b-8320-62a3b6a2bf48 0.032258063554763794\nQ1 Update 6797ade9-2418-4487-a3ee-bcebb62b9cf9 + 0.01587301678955555\n","stderr":"","success":true}' + role: tool + tool_call_id: call_dwvve4zt + - content: null + reasoning: We have Q3, Q2, Q1. Need sentiment classify each. Use llm on content. get_chunk each. Let's get chunks. + role: assistant + tool_calls: + - function: + arguments: '{"code":"chunks = []\nfor r in results:\n chunk = await get_chunk(r[''chunk_id''])\n chunks.append(chunk)\n print(chunk[''document_title''], + len(chunk[''content'']))\n"}' + name: execute_code + id: call_juxb6n6o + type: function + - content: '{"code":"chunks = []\nfor r in results:\n chunk = await get_chunk(r[''chunk_id''])\n chunks.append(chunk)\n print(chunk[''document_title''], + len(chunk[''content'']))\n","stdout":"","stderr":"NameError: name ''results'' is not defined","success":false}' + role: tool + tool_call_id: call_juxb6n6o + - content: null + reasoning: Need to keep results across calls? Variables persist? Might need to re-search. Let's re-execute. + role: assistant + tool_calls: + - function: + arguments: '{"code":"results = await search(\"quarterly update\", limit=20)\nchunks = []\nfor r in results:\n chunk + = await get_chunk(r[''chunk_id''])\n chunks.append(chunk)\n print(chunk[''document_title''], len(chunk[''content'']))\n"}' + name: execute_code + id: call_7m5eb6ho + type: function + - content: '{"code":"results = await search(\"quarterly update\", limit=20)\nchunks = []\nfor r in results:\n chunk + = await get_chunk(r[''chunk_id''])\n chunks.append(chunk)\n print(chunk[''document_title''], len(chunk[''content'']))\n","stdout":"Q3 + Update 168\nQ2 Update 161\nQ1 Update 152\n","stderr":"","success":true}' + role: tool + tool_call_id: call_7m5eb6ho + model: gpt-oss + reasoning_effort: low + stream: false + tool_choice: auto + tools: + - function: + description: |- + Execute Python code in a sandboxed interpreter. + + The code has access to haiku.rag functions (search, list_documents, + get_document, get_chunk, llm). + + Use print() to output results. + + Structured result with success status, stdout, and stderr. + + name: execute_code + parameters: + additionalProperties: false + properties: + code: + description: Python code to execute. + type: string + required: + - code + type: object + strict: true + type: function + - function: + description: Result from RLM agent execution. + name: final_result + parameters: + additionalProperties: false + properties: + answer: + description: The answer to the user's question + type: string + program: + description: The final consolidated program + type: string + required: + - answer + - program + type: object + strict: true + type: function + uri: http://localhost:11434/v1/chat/completions + response: + headers: + content-length: + - '783' + content-type: + - application/json + parsed_body: + choices: + - finish_reason: tool_calls + index: 0 + message: + content: '' + reasoning: Now classify sentiment. + role: assistant + tool_calls: + - function: + arguments: '{"code":"sentiments = {}\nfor chunk in chunks:\n prompt = f\"Classify the overall sentiment of + the following quarterly update as positive, negative, or mixed: {chunk[''content'']}\"\n resp = await llm(prompt)\n sentiments[chunk[''document_title'']] + = resp.strip()\n print(chunk[''document_title''], resp.strip())\n"}' + name: execute_code + id: call_g7dz17hi + index: 0 + type: function + created: 1771924537 + id: chatcmpl-853 + model: gpt-oss + object: chat.completion + system_fingerprint: fp_ollama + usage: + completion_tokens: 102 + prompt_tokens: 2278 + total_tokens: 2380 + status: + code: 200 + message: OK +- request: + headers: + accept: + - application/json + accept-encoding: + - gzip, deflate, zstd + connection: + - keep-alive + content-length: + - '10969' + content-type: + - application/json + host: + - localhost:11434 + method: POST + parsed_body: + messages: + - content: |- + You are a Recursive Language Model (RLM) agent that solves complex research questions by writing and executing Python code. + + IMPORTANT: You MUST use the `execute_code` tool to run Python code. The functions described below are ONLY available inside the execute_code tool - you cannot access them any other way. Always execute code to answer questions; do not just describe what code would do. + + CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just call them with `await`: + - results = await search("query") ✓ CORRECT + - from haiku.rag import search ✗ WRONG - will fail + - results = search("query") ✗ WRONG - must use await + + You have access to a sandboxed Python interpreter with these haiku.rag functions (use them directly with `await`, no imports needed): + + ## Available Functions + + ### await search(query, limit=10) -> list[dict] + Search the knowledge base using hybrid search (vector + full-text). + Returns list of dicts with keys: chunk_id, content, document_id, document_title, document_uri, score, page_numbers, headings + + ### await list_documents(limit=10, offset=0) -> list[dict] + List available documents in the knowledge base. + Returns list of dicts with keys: id, title, uri, created_at + + ### await get_document(id_or_title) -> str | None + Get the full text content of a document by ID, title, or URI. + Returns the document content as a string, or None if not found. + + ### await get_chunk(chunk_id) -> dict | None + Get a specific chunk by its ID (from search results). + Returns dict with keys: chunk_id, content, document_id, document_title, headings, page_numbers, labels + Use this to retrieve full chunk details and metadata for citation. + + ### await llm(prompt) -> str + Call an LLM directly with the given prompt. Returns the response as a string. + Use this for classification, summarization, extraction, or any task where you + already have the content and just need LLM reasoning. + + ## Pre-loaded Documents Variable + + If documents were pre-loaded for this session, a `documents` variable is available: + ```python + # documents is a list of dicts with keys: id, title, uri, content + for doc in documents: + print(doc['title'], len(doc['content'])) + ``` + Check if it exists with: `if 'documents' in dir(): ...` + + ## Available Python Features + + The interpreter supports: variables, arithmetic, strings, f-strings, lists, dicts, tuples, sets, loops, conditionals, comprehensions, functions, async/await, `map()`, `sorted()`/`.sort(key=...)`, try/except, and the `json` module. + + Not supported: imports (other than `json`), class definitions, generators/yield, match statements, decorators, `with` statements. + + For pattern matching or text extraction, use string methods (`str.split`, `str.find`, `str.startswith`, `in` operator) or the `llm()` function. + + ## Strategy Guide + + 1. **Explore First**: Start by listing documents or searching to understand what's available. Document names may differ from filenames (e.g., "tbmed593.pdf" might be stored as "TB MED 593" or similar). + 2. **If get_document returns None**: Use `await list_documents()` to see actual document titles, or `await search()` to find relevant content. + 3. **Iterative Refinement**: Run code, examine results, adjust your approach based on what you find. + 4. **Use print() Liberally**: The sandbox captures stdout - print intermediate results to see what you're working with. + 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and data structures. + 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use `await llm()`. + 7. **Cite Your Sources**: Use get_chunk() to retrieve chunk metadata for citations. Track which documents/chunks informed your answer. + + ## Example Patterns + + ### Counting documents matching a condition + ```python + docs = await list_documents(limit=100) + count = 0 + for doc in docs: + content = await get_document(doc['id']) + if content and 'keyword' in content.lower(): + count += 1 + print(f"Found in: {doc['title']}") + print(f"Total: {count}") + ``` + + ### Extracting data with llm() + ```python + numbers = [] + results = await search("financial data", limit=20) + for r in results: + extracted = await llm(f"Extract all dollar amounts from this text as a comma-separated list of numbers (no $ signs): {r['content']}") + for part in extracted.split(','): + part = part.strip().replace(',', '') + if part.isdigit(): + numbers.append(int(part)) + if numbers: + print(f"Average: {sum(numbers) / len(numbers)}") + ``` + + ### Using search results with get_chunk for citations + ```python + results = await search("safety requirements", limit=5) + for r in results: + chunk = await get_chunk(r['chunk_id']) + print(f"From '{chunk['document_title']}', page {chunk['page_numbers']}: {chunk['content'][:100]}") + ``` + + ### Using llm() for classification + ```python + content = await get_document("Q1 Report") + sentiment = await llm(f"Classify the sentiment as positive, negative, or mixed: {content}") + print(sentiment) + ``` + + ## Workflow + + 1. **ALWAYS start by using execute_code** to explore the knowledge base + 2. Run multiple code blocks as needed to gather information + 3. After collecting data, provide your final answer + + ## Output Format + + CRITICAL: Your final response MUST be valid JSON matching this exact schema: + ```json + {"answer": "Your complete answer here as a string", "program": "Your final consolidated program here as a string"} + ``` + + - `answer`: A clear answer to the user's question with key findings and references to specific documents/chunks. + - `program`: A single, self-contained Python program that produces the answer. Consolidate your exploratory code executions into one clean script. + + Do NOT return arbitrary JSON structures. Always use the exact format: {"answer": "...", "program": "..."} + + CRITICAL: You MUST call execute_code at least once before providing your answer. Never give up without trying to execute code first. + role: system + - content: Analyze the sentiment of each quarterly update. How many quarters were positive, negative, and mixed? + role: user + - content: null + reasoning: Need to search for quarterly updates. Let's search "quarterly update". + role: assistant + tool_calls: + - function: + arguments: '{"code":"results = await search(\"quarterly update\", limit=20)\nprint(len(results))\nfor r in results[:5]:\n print(r[''document_title''], + r[''chunk_id''], r[''score''])\n"}' + name: execute_code + id: call_dwvve4zt + type: function + - content: '{"code":"results = await search(\"quarterly update\", limit=20)\nprint(len(results))\nfor r in results[:5]:\n print(r[''document_title''], + r[''chunk_id''], r[''score''])\n","stdout":"3\nQ3 Update d651d341-7708-4ff3-9d4b-54bcd41694bf 0.032786883413791656\nQ2 + Update f7a3ff56-4f07-479b-8320-62a3b6a2bf48 0.032258063554763794\nQ1 Update 6797ade9-2418-4487-a3ee-bcebb62b9cf9 + 0.01587301678955555\n","stderr":"","success":true}' + role: tool + tool_call_id: call_dwvve4zt + - content: null + reasoning: We have Q3, Q2, Q1. Need sentiment classify each. Use llm on content. get_chunk each. Let's get chunks. + role: assistant + tool_calls: + - function: + arguments: '{"code":"chunks = []\nfor r in results:\n chunk = await get_chunk(r[''chunk_id''])\n chunks.append(chunk)\n print(chunk[''document_title''], + len(chunk[''content'']))\n"}' + name: execute_code + id: call_juxb6n6o + type: function + - content: '{"code":"chunks = []\nfor r in results:\n chunk = await get_chunk(r[''chunk_id''])\n chunks.append(chunk)\n print(chunk[''document_title''], + len(chunk[''content'']))\n","stdout":"","stderr":"NameError: name ''results'' is not defined","success":false}' + role: tool + tool_call_id: call_juxb6n6o + - content: null + reasoning: Need to keep results across calls? Variables persist? Might need to re-search. Let's re-execute. + role: assistant + tool_calls: + - function: + arguments: '{"code":"results = await search(\"quarterly update\", limit=20)\nchunks = []\nfor r in results:\n chunk + = await get_chunk(r[''chunk_id''])\n chunks.append(chunk)\n print(chunk[''document_title''], len(chunk[''content'']))\n"}' + name: execute_code + id: call_7m5eb6ho + type: function + - content: '{"code":"results = await search(\"quarterly update\", limit=20)\nchunks = []\nfor r in results:\n chunk + = await get_chunk(r[''chunk_id''])\n chunks.append(chunk)\n print(chunk[''document_title''], len(chunk[''content'']))\n","stdout":"Q3 + Update 168\nQ2 Update 161\nQ1 Update 152\n","stderr":"","success":true}' + role: tool + tool_call_id: call_7m5eb6ho + - content: null + reasoning: Now classify sentiment. + role: assistant + tool_calls: + - function: + arguments: '{"code":"sentiments = {}\nfor chunk in chunks:\n prompt = f\"Classify the overall sentiment of + the following quarterly update as positive, negative, or mixed: {chunk[''content'']}\"\n resp = await llm(prompt)\n sentiments[chunk[''document_title'']] + = resp.strip()\n print(chunk[''document_title''], resp.strip())\n"}' + name: execute_code + id: call_g7dz17hi + type: function + - content: '{"code":"sentiments = {}\nfor chunk in chunks:\n prompt = f\"Classify the overall sentiment of the following + quarterly update as positive, negative, or mixed: {chunk[''content'']}\"\n resp = await llm(prompt)\n sentiments[chunk[''document_title'']] + = resp.strip()\n print(chunk[''document_title''], resp.strip())\n","stdout":"","stderr":"NameError: name ''chunks'' + is not defined","success":false}' + role: tool + tool_call_id: call_g7dz17hi + model: gpt-oss + reasoning_effort: low + stream: false + tool_choice: auto + tools: + - function: + description: |- + Execute Python code in a sandboxed interpreter. + + The code has access to haiku.rag functions (search, list_documents, + get_document, get_chunk, llm). + + Use print() to output results. + + Structured result with success status, stdout, and stderr. + + name: execute_code + parameters: + additionalProperties: false + properties: + code: + description: Python code to execute. + type: string + required: + - code + type: object + strict: true + type: function + - function: + description: Result from RLM agent execution. + name: final_result + parameters: + additionalProperties: false + properties: + answer: + description: The answer to the user's question + type: string + program: + description: The final consolidated program + type: string + required: + - answer + - program + type: object + strict: true + type: function + uri: http://localhost:11434/v1/chat/completions + response: + headers: + content-length: + - '1045' + content-type: + - application/json + parsed_body: + choices: + - finish_reason: tool_calls + index: 0 + message: + content: '' + reasoning: We need to re-define chunks. Let's re-run entire block. + role: assistant + tool_calls: + - function: + arguments: '{"code":"results = await search(\"quarterly update\", limit=20)\nchunks = []\nfor r in results:\n chunk + = await get_chunk(r[''chunk_id''])\n chunks.append(chunk)\n print(chunk[''document_title''], len(chunk[''content'']))\n\nsentiments + = {}\nfor chunk in chunks:\n prompt = f\"Classify the overall sentiment of the following quarterly update + as positive, negative, or mixed: {chunk[''content'']}\"\n resp = await llm(prompt)\n sentiments[chunk[''document_title'']] + = resp.strip()\n print(chunk[''document_title''], resp.strip())\n"}' + name: execute_code + id: call_knskt3jd + index: 0 + type: function + created: 1771924541 + id: chatcmpl-263 + model: gpt-oss + object: chat.completion + system_fingerprint: fp_ollama + usage: + completion_tokens: 171 + prompt_tokens: 2489 + total_tokens: 2660 + status: + code: 200 + message: OK +- request: + headers: + accept: + - application/json + accept-encoding: + - gzip, deflate, zstd + connection: + - keep-alive + content-length: + - '86' + content-type: + - application/json + host: + - localhost:11434 + method: POST + parsed_body: + encoding_format: base64 + input: + - quarterly update + model: qwen3-embedding:4b + uri: http://localhost:11434/v1/embeddings + response: + headers: + content-type: + - application/json + transfer-encoding: + - chunked + parsed_body: + data: + - embedding: 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 + index: 0 + object: embedding + model: qwen3-embedding:4b + object: list + usage: + prompt_tokens: 4 + total_tokens: 4 + status: + code: 200 + message: OK +- request: + headers: + accept: + - application/json + accept-encoding: + - gzip, deflate, zstd + connection: + - keep-alive + content-length: + - '367' + content-type: + - application/json + host: + - localhost:11434 + method: POST + parsed_body: + messages: + - content: 'Classify the overall sentiment of the following quarterly update as positive, negative, or mixed: Mixed + results this quarter. While product quality improved, marketing campaigns underperformed. Revenue was flat compared + to last year but customer retention increased.' + role: user + model: gpt-oss + reasoning_effort: low + stream: false + uri: http://localhost:11434/v1/chat/completions + response: + headers: + content-length: + - '393' content-type: - application/json parsed_body: @@ -635,25 +1570,434 @@ interactions: - finish_reason: stop index: 0 message: - content: '{"answer":"Among the quarterly updates examined, there is one positive quarter (Q1), one negative quarter - (Q2), and one mixed quarter (Q3).","program":"# Find quarterly update chunks\nresults = search(\"quarterly update\", - limit=50)\n# Classify sentiment via llm\nsentiment_counts = {\"positive\":0, \"negative\":0, \"mixed\":0}\nfor - res in results:\n content = res[''content'']\n sentiment = llm(f\"Classify the sentiment of the following - corporate update as ''positive'', ''negative'', or ''mixed'':\\n\\n{content}\")\n # Normalize output\n sentiment - = sentiment.lower().split()[0]\n if sentiment in sentiment_counts:\n sentiment_counts[sentiment] += - 1\n else:\n sentiment_counts[''mixed''] += 1\nprint(sentiment_counts)\n# Output the counts\nprint(f\"Positive: - {sentiment_counts[''positive'']}\\nNegative: {sentiment_counts[''negative'']}\\nMixed: {sentiment_counts[''mixed'']}\")"}' - reasoning: We have 3 quarterly updates Q1 positive, Q2 negative, Q3 mixed. Need count. + content: '**Overall sentiment:** Mixed.' + reasoning: 'We need to classify overall sentiment: mixed. Probably "mixed".' role: assistant - created: 1770373411 - id: chatcmpl-392 + created: 1771924541 + id: chatcmpl-966 model: gpt-oss object: chat.completion system_fingerprint: fp_ollama usage: - completion_tokens: 264 - prompt_tokens: 2257 - total_tokens: 2521 + completion_tokens: 29 + prompt_tokens: 112 + total_tokens: 141 + status: + code: 200 + message: OK +- request: + headers: + accept: + - application/json + accept-encoding: + - gzip, deflate, zstd + connection: + - keep-alive + content-length: + - '360' + content-type: + - application/json + host: + - localhost:11434 + method: POST + parsed_body: + messages: + - content: 'Classify the overall sentiment of the following quarterly update as positive, negative, or mixed: We faced + significant challenges this quarter. Supply chain issues caused delays, and we missed our revenue target by 15%. + Several key employees left the company.' + role: user + model: gpt-oss + reasoning_effort: low + stream: false + uri: http://localhost:11434/v1/chat/completions + response: + headers: + content-length: + - '361' + content-type: + - application/json + parsed_body: + choices: + - finish_reason: stop + index: 0 + message: + content: '**Negative**' + reasoning: 'We must classify sentiment overall: it''s negative.' + role: assistant + created: 1771924542 + id: chatcmpl-538 + model: gpt-oss + object: chat.completion + system_fingerprint: fp_ollama + usage: + completion_tokens: 22 + prompt_tokens: 115 + total_tokens: 137 + status: + code: 200 + message: OK +- request: + headers: + accept: + - application/json + accept-encoding: + - gzip, deflate, zstd + connection: + - keep-alive + content-length: + - '351' + content-type: + - application/json + host: + - localhost:11434 + method: POST + parsed_body: + messages: + - content: 'Classify the overall sentiment of the following quarterly update as positive, negative, or mixed: The new + product launch exceeded expectations. Sales grew 40% and customer feedback has been overwhelmingly positive. Team + morale is at an all-time high.' + role: user + model: gpt-oss + reasoning_effort: low + stream: false + uri: http://localhost:11434/v1/chat/completions + response: + headers: + content-length: + - '320' + content-type: + - application/json + parsed_body: + choices: + - finish_reason: stop + index: 0 + message: + content: Positive + reasoning: All positive. + role: assistant + created: 1771924543 + id: chatcmpl-177 + model: gpt-oss + object: chat.completion + system_fingerprint: fp_ollama + usage: + completion_tokens: 14 + prompt_tokens: 114 + total_tokens: 128 + status: + code: 200 + message: OK +- request: + headers: + accept: + - application/json + accept-encoding: + - gzip, deflate, zstd + connection: + - keep-alive + content-length: + - '12558' + content-type: + - application/json + host: + - localhost:11434 + method: POST + parsed_body: + messages: + - content: |- + You are a Recursive Language Model (RLM) agent that solves complex research questions by writing and executing Python code. + + IMPORTANT: You MUST use the `execute_code` tool to run Python code. The functions described below are ONLY available inside the execute_code tool - you cannot access them any other way. Always execute code to answer questions; do not just describe what code would do. + + CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just call them with `await`: + - results = await search("query") ✓ CORRECT + - from haiku.rag import search ✗ WRONG - will fail + - results = search("query") ✗ WRONG - must use await + + You have access to a sandboxed Python interpreter with these haiku.rag functions (use them directly with `await`, no imports needed): + + ## Available Functions + + ### await search(query, limit=10) -> list[dict] + Search the knowledge base using hybrid search (vector + full-text). + Returns list of dicts with keys: chunk_id, content, document_id, document_title, document_uri, score, page_numbers, headings + + ### await list_documents(limit=10, offset=0) -> list[dict] + List available documents in the knowledge base. + Returns list of dicts with keys: id, title, uri, created_at + + ### await get_document(id_or_title) -> str | None + Get the full text content of a document by ID, title, or URI. + Returns the document content as a string, or None if not found. + + ### await get_chunk(chunk_id) -> dict | None + Get a specific chunk by its ID (from search results). + Returns dict with keys: chunk_id, content, document_id, document_title, headings, page_numbers, labels + Use this to retrieve full chunk details and metadata for citation. + + ### await llm(prompt) -> str + Call an LLM directly with the given prompt. Returns the response as a string. + Use this for classification, summarization, extraction, or any task where you + already have the content and just need LLM reasoning. + + ## Pre-loaded Documents Variable + + If documents were pre-loaded for this session, a `documents` variable is available: + ```python + # documents is a list of dicts with keys: id, title, uri, content + for doc in documents: + print(doc['title'], len(doc['content'])) + ``` + Check if it exists with: `if 'documents' in dir(): ...` + + ## Available Python Features + + The interpreter supports: variables, arithmetic, strings, f-strings, lists, dicts, tuples, sets, loops, conditionals, comprehensions, functions, async/await, `map()`, `sorted()`/`.sort(key=...)`, try/except, and the `json` module. + + Not supported: imports (other than `json`), class definitions, generators/yield, match statements, decorators, `with` statements. + + For pattern matching or text extraction, use string methods (`str.split`, `str.find`, `str.startswith`, `in` operator) or the `llm()` function. + + ## Strategy Guide + + 1. **Explore First**: Start by listing documents or searching to understand what's available. Document names may differ from filenames (e.g., "tbmed593.pdf" might be stored as "TB MED 593" or similar). + 2. **If get_document returns None**: Use `await list_documents()` to see actual document titles, or `await search()` to find relevant content. + 3. **Iterative Refinement**: Run code, examine results, adjust your approach based on what you find. + 4. **Use print() Liberally**: The sandbox captures stdout - print intermediate results to see what you're working with. + 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and data structures. + 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use `await llm()`. + 7. **Cite Your Sources**: Use get_chunk() to retrieve chunk metadata for citations. Track which documents/chunks informed your answer. + + ## Example Patterns + + ### Counting documents matching a condition + ```python + docs = await list_documents(limit=100) + count = 0 + for doc in docs: + content = await get_document(doc['id']) + if content and 'keyword' in content.lower(): + count += 1 + print(f"Found in: {doc['title']}") + print(f"Total: {count}") + ``` + + ### Extracting data with llm() + ```python + numbers = [] + results = await search("financial data", limit=20) + for r in results: + extracted = await llm(f"Extract all dollar amounts from this text as a comma-separated list of numbers (no $ signs): {r['content']}") + for part in extracted.split(','): + part = part.strip().replace(',', '') + if part.isdigit(): + numbers.append(int(part)) + if numbers: + print(f"Average: {sum(numbers) / len(numbers)}") + ``` + + ### Using search results with get_chunk for citations + ```python + results = await search("safety requirements", limit=5) + for r in results: + chunk = await get_chunk(r['chunk_id']) + print(f"From '{chunk['document_title']}', page {chunk['page_numbers']}: {chunk['content'][:100]}") + ``` + + ### Using llm() for classification + ```python + content = await get_document("Q1 Report") + sentiment = await llm(f"Classify the sentiment as positive, negative, or mixed: {content}") + print(sentiment) + ``` + + ## Workflow + + 1. **ALWAYS start by using execute_code** to explore the knowledge base + 2. Run multiple code blocks as needed to gather information + 3. After collecting data, provide your final answer + + ## Output Format + + CRITICAL: Your final response MUST be valid JSON matching this exact schema: + ```json + {"answer": "Your complete answer here as a string", "program": "Your final consolidated program here as a string"} + ``` + + - `answer`: A clear answer to the user's question with key findings and references to specific documents/chunks. + - `program`: A single, self-contained Python program that produces the answer. Consolidate your exploratory code executions into one clean script. + + Do NOT return arbitrary JSON structures. Always use the exact format: {"answer": "...", "program": "..."} + + CRITICAL: You MUST call execute_code at least once before providing your answer. Never give up without trying to execute code first. + role: system + - content: Analyze the sentiment of each quarterly update. How many quarters were positive, negative, and mixed? + role: user + - content: null + reasoning: Need to search for quarterly updates. Let's search "quarterly update". + role: assistant + tool_calls: + - function: + arguments: '{"code":"results = await search(\"quarterly update\", limit=20)\nprint(len(results))\nfor r in results[:5]:\n print(r[''document_title''], + r[''chunk_id''], r[''score''])\n"}' + name: execute_code + id: call_dwvve4zt + type: function + - content: '{"code":"results = await search(\"quarterly update\", limit=20)\nprint(len(results))\nfor r in results[:5]:\n print(r[''document_title''], + r[''chunk_id''], r[''score''])\n","stdout":"3\nQ3 Update d651d341-7708-4ff3-9d4b-54bcd41694bf 0.032786883413791656\nQ2 + Update f7a3ff56-4f07-479b-8320-62a3b6a2bf48 0.032258063554763794\nQ1 Update 6797ade9-2418-4487-a3ee-bcebb62b9cf9 + 0.01587301678955555\n","stderr":"","success":true}' + role: tool + tool_call_id: call_dwvve4zt + - content: null + reasoning: We have Q3, Q2, Q1. Need sentiment classify each. Use llm on content. get_chunk each. Let's get chunks. + role: assistant + tool_calls: + - function: + arguments: '{"code":"chunks = []\nfor r in results:\n chunk = await get_chunk(r[''chunk_id''])\n chunks.append(chunk)\n print(chunk[''document_title''], + len(chunk[''content'']))\n"}' + name: execute_code + id: call_juxb6n6o + type: function + - content: '{"code":"chunks = []\nfor r in results:\n chunk = await get_chunk(r[''chunk_id''])\n chunks.append(chunk)\n print(chunk[''document_title''], + len(chunk[''content'']))\n","stdout":"","stderr":"NameError: name ''results'' is not defined","success":false}' + role: tool + tool_call_id: call_juxb6n6o + - content: null + reasoning: Need to keep results across calls? Variables persist? Might need to re-search. Let's re-execute. + role: assistant + tool_calls: + - function: + arguments: '{"code":"results = await search(\"quarterly update\", limit=20)\nchunks = []\nfor r in results:\n chunk + = await get_chunk(r[''chunk_id''])\n chunks.append(chunk)\n print(chunk[''document_title''], len(chunk[''content'']))\n"}' + name: execute_code + id: call_7m5eb6ho + type: function + - content: '{"code":"results = await search(\"quarterly update\", limit=20)\nchunks = []\nfor r in results:\n chunk + = await get_chunk(r[''chunk_id''])\n chunks.append(chunk)\n print(chunk[''document_title''], len(chunk[''content'']))\n","stdout":"Q3 + Update 168\nQ2 Update 161\nQ1 Update 152\n","stderr":"","success":true}' + role: tool + tool_call_id: call_7m5eb6ho + - content: null + reasoning: Now classify sentiment. + role: assistant + tool_calls: + - function: + arguments: '{"code":"sentiments = {}\nfor chunk in chunks:\n prompt = f\"Classify the overall sentiment of + the following quarterly update as positive, negative, or mixed: {chunk[''content'']}\"\n resp = await llm(prompt)\n sentiments[chunk[''document_title'']] + = resp.strip()\n print(chunk[''document_title''], resp.strip())\n"}' + name: execute_code + id: call_g7dz17hi + type: function + - content: '{"code":"sentiments = {}\nfor chunk in chunks:\n prompt = f\"Classify the overall sentiment of the following + quarterly update as positive, negative, or mixed: {chunk[''content'']}\"\n resp = await llm(prompt)\n sentiments[chunk[''document_title'']] + = resp.strip()\n print(chunk[''document_title''], resp.strip())\n","stdout":"","stderr":"NameError: name ''chunks'' + is not defined","success":false}' + role: tool + tool_call_id: call_g7dz17hi + - content: null + reasoning: We need to re-define chunks. Let's re-run entire block. + role: assistant + tool_calls: + - function: + arguments: '{"code":"results = await search(\"quarterly update\", limit=20)\nchunks = []\nfor r in results:\n chunk + = await get_chunk(r[''chunk_id''])\n chunks.append(chunk)\n print(chunk[''document_title''], len(chunk[''content'']))\n\nsentiments + = {}\nfor chunk in chunks:\n prompt = f\"Classify the overall sentiment of the following quarterly update + as positive, negative, or mixed: {chunk[''content'']}\"\n resp = await llm(prompt)\n sentiments[chunk[''document_title'']] + = resp.strip()\n print(chunk[''document_title''], resp.strip())\n"}' + name: execute_code + id: call_knskt3jd + type: function + - content: '{"code":"results = await search(\"quarterly update\", limit=20)\nchunks = []\nfor r in results:\n chunk + = await get_chunk(r[''chunk_id''])\n chunks.append(chunk)\n print(chunk[''document_title''], len(chunk[''content'']))\n\nsentiments + = {}\nfor chunk in chunks:\n prompt = f\"Classify the overall sentiment of the following quarterly update as + positive, negative, or mixed: {chunk[''content'']}\"\n resp = await llm(prompt)\n sentiments[chunk[''document_title'']] + = resp.strip()\n print(chunk[''document_title''], resp.strip())\n","stdout":"Q3 Update 168\nQ2 Update 161\nQ1 + Update 152\nQ3 Update **Overall sentiment:** Mixed.\nQ2 Update **Negative**\nQ1 Update Positive\n","stderr":"","success":true}' + role: tool + tool_call_id: call_knskt3jd + model: gpt-oss + reasoning_effort: low + stream: false + tool_choice: auto + tools: + - function: + description: |- + Execute Python code in a sandboxed interpreter. + + The code has access to haiku.rag functions (search, list_documents, + get_document, get_chunk, llm). + + Use print() to output results. + + Structured result with success status, stdout, and stderr. + + name: execute_code + parameters: + additionalProperties: false + properties: + code: + description: Python code to execute. + type: string + required: + - code + type: object + strict: true + type: function + - function: + description: Result from RLM agent execution. + name: final_result + parameters: + additionalProperties: false + properties: + answer: + description: The answer to the user's question + type: string + program: + description: The final consolidated program + type: string + required: + - answer + - program + type: object + strict: true + type: function + uri: http://localhost:11434/v1/chat/completions + response: + headers: + content-length: + - '1912' + content-type: + - application/json + parsed_body: + choices: + - finish_reason: stop + index: 0 + message: + content: '{"answer":"There are three quarterly updates in the knowledge base. After classifying each update, we + find one positive, one negative, and one mixed (neutral) sentiment.","program":"#!/usr/bin/env python3\nimport + json\n\n# RLM workflow to count sentiments of quarterly updates\n\nasync def main():\n # Search for quarterly + updates\n results = await search(\"quarterly update\", limit=20)\n # Retrieve chunk details for each result\n chunks + = []\n for r in results:\n chunk = await get_chunk(r[''chunk_id''])\n chunks.append(chunk)\n # + Classify sentiment of each chunk using the LLM\n sentiments = {}\n for chunk in chunks:\n prompt + = (\n f\"Classify the overall sentiment of the following quarterly update as positive, negative, or + mixed: \"\n f\"{chunk[''content'']}\"\n )\n resp = await llm(prompt)\n sentiments[chunk[''document_title'']] + = resp.strip()\n # Count occurrences of each sentiment category\n counts = {\"positive\":0, \"negative\":0, + \"mixed\":0}\n for s in sentiments.values():\n key = s.lower()\n if key in counts:\n counts[key] + += 1\n else:\n # handle unexpected output\n pass\n # Output results\n print(\"Sentiment + counts:\")\n print(json.dumps(counts, indent=2))\n\n# Run the asynchronous main function\nif __name__ == \"__main__\":\n import + asyncio\n asyncio.run(main())\n"}' + reasoning: 'We have counts: Mixed 1, Negative 1, Positive 1. Provide answer and program.' + role: assistant + created: 1771924554 + id: chatcmpl-522 + model: gpt-oss + object: chat.completion + system_fingerprint: fp_ollama + usage: + completion_tokens: 386 + prompt_tokens: 2859 + total_tokens: 3245 status: code: 200 message: OK diff --git a/tests/cassettes/test_rlm/TestClientRLMIntegration.test_rlm_with_filter.yaml b/tests/cassettes/test_rlm/TestClientRLMIntegration.test_rlm_with_filter.yaml index 0b23f8f5..3a9012b8 100644 --- a/tests/cassettes/test_rlm/TestClientRLMIntegration.test_rlm_with_filter.yaml +++ b/tests/cassettes/test_rlm/TestClientRLMIntegration.test_rlm_with_filter.yaml @@ -128,7 +128,7 @@ interactions: connection: - keep-alive content-length: - - '7768' + - '7319' content-type: - application/json host: @@ -141,32 +141,33 @@ interactions: IMPORTANT: You MUST use the `execute_code` tool to run Python code. The functions described below are ONLY available inside the execute_code tool - you cannot access them any other way. Always execute code to answer questions; do not just describe what code would do. - CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just use them directly: - - search("query") ✓ CORRECT + CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just call them with `await`: + - results = await search("query") ✓ CORRECT - from haiku.rag import search ✗ WRONG - will fail + - results = search("query") ✗ WRONG - must use await - You have access to a sandboxed Python environment with these haiku.rag functions (use them directly, no imports needed): + You have access to a sandboxed Python interpreter with these haiku.rag functions (use them directly with `await`, no imports needed): ## Available Functions - ### search(query, limit=10) -> list[dict] + ### await search(query, limit=10) -> list[dict] Search the knowledge base using hybrid search (vector + full-text). Returns list of dicts with keys: chunk_id, content, document_id, document_title, document_uri, score, page_numbers, headings - ### list_documents(limit=10, offset=0) -> list[dict] + ### await list_documents(limit=10, offset=0) -> list[dict] List available documents in the knowledge base. Returns list of dicts with keys: id, title, uri, created_at - ### get_document(id_or_title) -> str | None + ### await get_document(id_or_title) -> str | None Get the full text content of a document by ID, title, or URI. Returns the document content as a string, or None if not found. - ### get_docling_document(id_or_title) -> DoclingDocument | None - Get the structured DoclingDocument object for advanced analysis. - Returns a DoclingDocument object, or None if not found. - See "DoclingDocument API" section below for how to use it. + ### await get_chunk(chunk_id) -> dict | None + Get a specific chunk by its ID (from search results). + Returns dict with keys: chunk_id, content, document_id, document_title, headings, page_numbers, labels + Use this to retrieve full chunk details and metadata for citation. - ### llm(prompt) -> str + ### await llm(prompt) -> str Call an LLM directly with the given prompt. Returns the response as a string. Use this for classification, summarization, extraction, or any task where you already have the content and just need LLM reasoning. @@ -181,93 +182,64 @@ interactions: ``` Check if it exists with: `if 'documents' in dir(): ...` - ## Standard Library Modules - You can import any Python standard library module. + ## Available Python Features + + The interpreter supports: variables, arithmetic, strings, f-strings, lists, dicts, tuples, sets, loops, conditionals, comprehensions, functions, async/await, `map()`, `sorted()`/`.sort(key=...)`, try/except, and the `json` module. + + Not supported: imports (other than `json`), class definitions, generators/yield, match statements, decorators, `with` statements. + + For pattern matching or text extraction, use string methods (`str.split`, `str.find`, `str.startswith`, `in` operator) or the `llm()` function. ## Strategy Guide 1. **Explore First**: Start by listing documents or searching to understand what's available. Document names may differ from filenames (e.g., "tbmed593.pdf" might be stored as "TB MED 593" or similar). - 2. **If get_document returns None**: Use `list_documents()` to see actual document titles, or `search()` to find relevant content. + 2. **If get_document returns None**: Use `await list_documents()` to see actual document titles, or `await search()` to find relevant content. 3. **Iterative Refinement**: Run code, examine results, adjust your approach based on what you find. - 4. **Use print() Liberally**: The REPL captures stdout - print intermediate results to see what you're working with. - 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and use collections. - 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use llm(). - 7. **Cite Your Sources**: Track which documents/chunks informed your answer for citation. - - ## DoclingDocument API - - When you call `get_docling_document(id_or_title)`, you get a DoclingDocument object for structured document analysis. - - ### Properties - - `doc.texts` - List of all text items (paragraphs, headings, etc.) - - `doc.tables` - List of all tables - - `doc.pictures` - List of all pictures/figures - - `doc.name` - Document name - - ### Methods - - `doc.iterate_items(with_groups=False)` - Iterate all items with hierarchy level - Returns tuples of (item, level) where level is nesting depth - - `doc.export_to_markdown()` - Export entire document as markdown string - - ### Text Item Properties - - `item.text` - The text content - - `item.label` - Type: title, paragraph, section_header, list_item, etc. (lowercase enum values) - - `item.prov` - Provenance (page numbers, bounding boxes) - - ### Table Access - - `table.data.num_rows`, `table.data.num_cols` - Dimensions - - `table.data.table_cells` - List of TableCell objects - - `cell.text`, `cell.start_row_offset_idx`, `cell.start_col_offset_idx` - - ### Example Usage - ```python - doc = get_docling_document("My Document") - - # Get all headings - headings = [t.text for t in doc.texts if "header" in str(t.label)] - - # Iterate with structure - for item, level in doc.iterate_items(): - print(" " * level + item.text[:50]) - - # Extract table data - for table in doc.tables: - for cell in table.data.table_cells: - print(f"Row {cell.start_row_offset_idx}, Col {cell.start_col_offset_idx}: {cell.text}") - ``` + 4. **Use print() Liberally**: The sandbox captures stdout - print intermediate results to see what you're working with. + 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and data structures. + 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use `await llm()`. + 7. **Cite Your Sources**: Use get_chunk() to retrieve chunk metadata for citations. Track which documents/chunks informed your answer. ## Example Patterns ### Counting documents matching a condition ```python - docs = list_documents(limit=100) + docs = await list_documents(limit=100) count = 0 for doc in docs: - content = get_document(doc['id']) + content = await get_document(doc['id']) if content and 'keyword' in content.lower(): count += 1 print(f"Found in: {doc['title']}") print(f"Total: {count}") ``` - ### Aggregating data across documents + ### Extracting data with llm() ```python - import re numbers = [] - results = search("financial data", limit=20) + results = await search("financial data", limit=20) for r in results: - matches = re.findall(r'\$([\d,]+)', r['content']) - for m in matches: - numbers.append(int(m.replace(',', ''))) - print(f"Average: ${sum(numbers)/len(numbers):,.2f}") + extracted = await llm(f"Extract all dollar amounts from this text as a comma-separated list of numbers (no $ signs): {r['content']}") + for part in extracted.split(','): + part = part.strip().replace(',', '') + if part.isdigit(): + numbers.append(int(part)) + if numbers: + print(f"Average: {sum(numbers) / len(numbers)}") + ``` + + ### Using search results with get_chunk for citations + ```python + results = await search("safety requirements", limit=5) + for r in results: + chunk = await get_chunk(r['chunk_id']) + print(f"From '{chunk['document_title']}', page {chunk['page_numbers']}: {chunk['content'][:100]}") ``` ### Using llm() for classification ```python - # Get document content - content = get_document("Q1 Report") - # Use llm() to classify sentiment - sentiment = llm(f"Classify the sentiment as positive, negative, or mixed: {content}") + content = await get_document("Q1 Report") + sentiment = await llm(f"Classify the sentiment as positive, negative, or mixed: {content}") print(sentiment) ``` @@ -300,11 +272,10 @@ interactions: tools: - function: description: |- - Execute Python code in a Docker-sandboxed environment. + Execute Python code in a sandboxed interpreter. The code has access to haiku.rag functions (search, list_documents, - get_document, get_docling_document, llm) and any Python standard - library module. + get_document, get_chunk, llm). Use print() to output results. @@ -344,7 +315,7 @@ interactions: response: headers: content-length: - - '517' + - '547' content-type: - application/json parsed_body: @@ -353,24 +324,24 @@ interactions: index: 0 message: content: '' - reasoning: Need to get list_documents. + reasoning: Need to list documents. role: assistant tool_calls: - function: - arguments: '{"code":"docs = list_documents(limit=1000)\nprint(len(docs))"}' + arguments: '{"code":"docs = await list_documents(limit=1000, offset=0)\nprint(len(docs))\nprint(docs[:3])"}' name: execute_code - id: call_ly3bn3y0 + id: call_pu34e9fx index: 0 type: function - created: 1770373376 - id: chatcmpl-196 + created: 1771924517 + id: chatcmpl-236 model: gpt-oss object: chat.completion system_fingerprint: fp_ollama usage: - completion_tokens: 44 - prompt_tokens: 1745 - total_tokens: 1789 + completion_tokens: 56 + prompt_tokens: 1621 + total_tokens: 1677 status: code: 200 message: OK @@ -383,7 +354,7 @@ interactions: connection: - keep-alive content-length: - - '8201' + - '7939' content-type: - application/json host: @@ -396,32 +367,33 @@ interactions: IMPORTANT: You MUST use the `execute_code` tool to run Python code. The functions described below are ONLY available inside the execute_code tool - you cannot access them any other way. Always execute code to answer questions; do not just describe what code would do. - CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just use them directly: - - search("query") ✓ CORRECT + CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just call them with `await`: + - results = await search("query") ✓ CORRECT - from haiku.rag import search ✗ WRONG - will fail + - results = search("query") ✗ WRONG - must use await - You have access to a sandboxed Python environment with these haiku.rag functions (use them directly, no imports needed): + You have access to a sandboxed Python interpreter with these haiku.rag functions (use them directly with `await`, no imports needed): ## Available Functions - ### search(query, limit=10) -> list[dict] + ### await search(query, limit=10) -> list[dict] Search the knowledge base using hybrid search (vector + full-text). Returns list of dicts with keys: chunk_id, content, document_id, document_title, document_uri, score, page_numbers, headings - ### list_documents(limit=10, offset=0) -> list[dict] + ### await list_documents(limit=10, offset=0) -> list[dict] List available documents in the knowledge base. Returns list of dicts with keys: id, title, uri, created_at - ### get_document(id_or_title) -> str | None + ### await get_document(id_or_title) -> str | None Get the full text content of a document by ID, title, or URI. Returns the document content as a string, or None if not found. - ### get_docling_document(id_or_title) -> DoclingDocument | None - Get the structured DoclingDocument object for advanced analysis. - Returns a DoclingDocument object, or None if not found. - See "DoclingDocument API" section below for how to use it. + ### await get_chunk(chunk_id) -> dict | None + Get a specific chunk by its ID (from search results). + Returns dict with keys: chunk_id, content, document_id, document_title, headings, page_numbers, labels + Use this to retrieve full chunk details and metadata for citation. - ### llm(prompt) -> str + ### await llm(prompt) -> str Call an LLM directly with the given prompt. Returns the response as a string. Use this for classification, summarization, extraction, or any task where you already have the content and just need LLM reasoning. @@ -436,93 +408,64 @@ interactions: ``` Check if it exists with: `if 'documents' in dir(): ...` - ## Standard Library Modules - You can import any Python standard library module. + ## Available Python Features + + The interpreter supports: variables, arithmetic, strings, f-strings, lists, dicts, tuples, sets, loops, conditionals, comprehensions, functions, async/await, `map()`, `sorted()`/`.sort(key=...)`, try/except, and the `json` module. + + Not supported: imports (other than `json`), class definitions, generators/yield, match statements, decorators, `with` statements. + + For pattern matching or text extraction, use string methods (`str.split`, `str.find`, `str.startswith`, `in` operator) or the `llm()` function. ## Strategy Guide 1. **Explore First**: Start by listing documents or searching to understand what's available. Document names may differ from filenames (e.g., "tbmed593.pdf" might be stored as "TB MED 593" or similar). - 2. **If get_document returns None**: Use `list_documents()` to see actual document titles, or `search()` to find relevant content. + 2. **If get_document returns None**: Use `await list_documents()` to see actual document titles, or `await search()` to find relevant content. 3. **Iterative Refinement**: Run code, examine results, adjust your approach based on what you find. - 4. **Use print() Liberally**: The REPL captures stdout - print intermediate results to see what you're working with. - 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and use collections. - 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use llm(). - 7. **Cite Your Sources**: Track which documents/chunks informed your answer for citation. - - ## DoclingDocument API - - When you call `get_docling_document(id_or_title)`, you get a DoclingDocument object for structured document analysis. - - ### Properties - - `doc.texts` - List of all text items (paragraphs, headings, etc.) - - `doc.tables` - List of all tables - - `doc.pictures` - List of all pictures/figures - - `doc.name` - Document name - - ### Methods - - `doc.iterate_items(with_groups=False)` - Iterate all items with hierarchy level - Returns tuples of (item, level) where level is nesting depth - - `doc.export_to_markdown()` - Export entire document as markdown string - - ### Text Item Properties - - `item.text` - The text content - - `item.label` - Type: title, paragraph, section_header, list_item, etc. (lowercase enum values) - - `item.prov` - Provenance (page numbers, bounding boxes) - - ### Table Access - - `table.data.num_rows`, `table.data.num_cols` - Dimensions - - `table.data.table_cells` - List of TableCell objects - - `cell.text`, `cell.start_row_offset_idx`, `cell.start_col_offset_idx` - - ### Example Usage - ```python - doc = get_docling_document("My Document") - - # Get all headings - headings = [t.text for t in doc.texts if "header" in str(t.label)] - - # Iterate with structure - for item, level in doc.iterate_items(): - print(" " * level + item.text[:50]) - - # Extract table data - for table in doc.tables: - for cell in table.data.table_cells: - print(f"Row {cell.start_row_offset_idx}, Col {cell.start_col_offset_idx}: {cell.text}") - ``` + 4. **Use print() Liberally**: The sandbox captures stdout - print intermediate results to see what you're working with. + 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and data structures. + 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use `await llm()`. + 7. **Cite Your Sources**: Use get_chunk() to retrieve chunk metadata for citations. Track which documents/chunks informed your answer. ## Example Patterns ### Counting documents matching a condition ```python - docs = list_documents(limit=100) + docs = await list_documents(limit=100) count = 0 for doc in docs: - content = get_document(doc['id']) + content = await get_document(doc['id']) if content and 'keyword' in content.lower(): count += 1 print(f"Found in: {doc['title']}") print(f"Total: {count}") ``` - ### Aggregating data across documents + ### Extracting data with llm() ```python - import re numbers = [] - results = search("financial data", limit=20) + results = await search("financial data", limit=20) for r in results: - matches = re.findall(r'\$([\d,]+)', r['content']) - for m in matches: - numbers.append(int(m.replace(',', ''))) - print(f"Average: ${sum(numbers)/len(numbers):,.2f}") + extracted = await llm(f"Extract all dollar amounts from this text as a comma-separated list of numbers (no $ signs): {r['content']}") + for part in extracted.split(','): + part = part.strip().replace(',', '') + if part.isdigit(): + numbers.append(int(part)) + if numbers: + print(f"Average: {sum(numbers) / len(numbers)}") + ``` + + ### Using search results with get_chunk for citations + ```python + results = await search("safety requirements", limit=5) + for r in results: + chunk = await get_chunk(r['chunk_id']) + print(f"From '{chunk['document_title']}', page {chunk['page_numbers']}: {chunk['content'][:100]}") ``` ### Using llm() for classification ```python - # Get document content - content = get_document("Q1 Report") - # Use llm() to classify sentiment - sentiment = llm(f"Classify the sentiment as positive, negative, or mixed: {content}") + content = await get_document("Q1 Report") + sentiment = await llm(f"Classify the sentiment as positive, negative, or mixed: {content}") print(sentiment) ``` @@ -548,20 +491,19 @@ interactions: role: system - content: How many documents are available? role: user - - content: |- - - Need to get list_documents. - + - content: null + reasoning: Need to list documents. role: assistant tool_calls: - function: - arguments: '{"code":"docs = list_documents(limit=1000)\nprint(len(docs))"}' + arguments: '{"code":"docs = await list_documents(limit=1000, offset=0)\nprint(len(docs))\nprint(docs[:3])"}' name: execute_code - id: call_ly3bn3y0 + id: call_pu34e9fx type: function - - content: '{"code":"docs = list_documents(limit=1000)\nprint(len(docs))","stdout":"1\n","stderr":"","success":true}' + - content: '{"code":"docs = await list_documents(limit=1000, offset=0)\nprint(len(docs))\nprint(docs[:3])","stdout":"1\n[{''id'': + ''90fce4e5-0d68-4736-acaf-d075e3ed4822'', ''title'': ''Cats'', ''uri'': None, ''created_at'': ''2026-02-24 11:15:14.007278''}]\n","stderr":"","success":true}' role: tool - tool_call_id: call_ly3bn3y0 + tool_call_id: call_pu34e9fx model: gpt-oss reasoning_effort: low stream: false @@ -569,11 +511,10 @@ interactions: tools: - function: description: |- - Execute Python code in a Docker-sandboxed environment. + Execute Python code in a sandboxed interpreter. The code has access to haiku.rag functions (search, list_documents, - get_document, get_docling_document, llm) and any Python standard - library module. + get_document, get_chunk, llm). Use print() to output results. @@ -613,7 +554,7 @@ interactions: response: headers: content-length: - - '424' + - '704' content-type: - application/json parsed_body: @@ -621,17 +562,20 @@ interactions: - finish_reason: stop index: 0 message: - content: '{"answer":"There is 1 document available in the knowledge base.","program":"docs = list_documents(limit=1000)\nprint(len(docs))"}' + content: '{"answer":"There is 1 document available in the knowledge base.","program":"# List all documents\nimport + json\nasync def main():\n docs = await list_documents(limit=1000, offset=0)\n print(\"Number of documents:\", + len(docs))\n # Optional: print first few document titles for reference\n for d in docs[:5]:\n print(f\"ID: + {d[''id'']}, Title: {d[''title'']}\")\n\nawait main()"}' role: assistant - created: 1770373377 - id: chatcmpl-195 + created: 1771924519 + id: chatcmpl-487 model: gpt-oss object: chat.completion system_fingerprint: fp_ollama usage: - completion_tokens: 39 - prompt_tokens: 1836 - total_tokens: 1875 + completion_tokens: 117 + prompt_tokens: 1793 + total_tokens: 1910 status: code: 200 message: OK diff --git a/tests/cassettes/test_rlm/TestClientRLMIntegration.test_rlm_with_preloaded_documents.yaml b/tests/cassettes/test_rlm/TestClientRLMIntegration.test_rlm_with_preloaded_documents.yaml index 2b29eaaf..0a78fe8b 100644 --- a/tests/cassettes/test_rlm/TestClientRLMIntegration.test_rlm_with_preloaded_documents.yaml +++ b/tests/cassettes/test_rlm/TestClientRLMIntegration.test_rlm_with_preloaded_documents.yaml @@ -88,7 +88,7 @@ interactions: connection: - keep-alive content-length: - - '7839' + - '7390' content-type: - application/json host: @@ -101,32 +101,33 @@ interactions: IMPORTANT: You MUST use the `execute_code` tool to run Python code. The functions described below are ONLY available inside the execute_code tool - you cannot access them any other way. Always execute code to answer questions; do not just describe what code would do. - CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just use them directly: - - search("query") ✓ CORRECT + CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just call them with `await`: + - results = await search("query") ✓ CORRECT - from haiku.rag import search ✗ WRONG - will fail + - results = search("query") ✗ WRONG - must use await - You have access to a sandboxed Python environment with these haiku.rag functions (use them directly, no imports needed): + You have access to a sandboxed Python interpreter with these haiku.rag functions (use them directly with `await`, no imports needed): ## Available Functions - ### search(query, limit=10) -> list[dict] + ### await search(query, limit=10) -> list[dict] Search the knowledge base using hybrid search (vector + full-text). Returns list of dicts with keys: chunk_id, content, document_id, document_title, document_uri, score, page_numbers, headings - ### list_documents(limit=10, offset=0) -> list[dict] + ### await list_documents(limit=10, offset=0) -> list[dict] List available documents in the knowledge base. Returns list of dicts with keys: id, title, uri, created_at - ### get_document(id_or_title) -> str | None + ### await get_document(id_or_title) -> str | None Get the full text content of a document by ID, title, or URI. Returns the document content as a string, or None if not found. - ### get_docling_document(id_or_title) -> DoclingDocument | None - Get the structured DoclingDocument object for advanced analysis. - Returns a DoclingDocument object, or None if not found. - See "DoclingDocument API" section below for how to use it. + ### await get_chunk(chunk_id) -> dict | None + Get a specific chunk by its ID (from search results). + Returns dict with keys: chunk_id, content, document_id, document_title, headings, page_numbers, labels + Use this to retrieve full chunk details and metadata for citation. - ### llm(prompt) -> str + ### await llm(prompt) -> str Call an LLM directly with the given prompt. Returns the response as a string. Use this for classification, summarization, extraction, or any task where you already have the content and just need LLM reasoning. @@ -141,93 +142,64 @@ interactions: ``` Check if it exists with: `if 'documents' in dir(): ...` - ## Standard Library Modules - You can import any Python standard library module. + ## Available Python Features + + The interpreter supports: variables, arithmetic, strings, f-strings, lists, dicts, tuples, sets, loops, conditionals, comprehensions, functions, async/await, `map()`, `sorted()`/`.sort(key=...)`, try/except, and the `json` module. + + Not supported: imports (other than `json`), class definitions, generators/yield, match statements, decorators, `with` statements. + + For pattern matching or text extraction, use string methods (`str.split`, `str.find`, `str.startswith`, `in` operator) or the `llm()` function. ## Strategy Guide 1. **Explore First**: Start by listing documents or searching to understand what's available. Document names may differ from filenames (e.g., "tbmed593.pdf" might be stored as "TB MED 593" or similar). - 2. **If get_document returns None**: Use `list_documents()` to see actual document titles, or `search()` to find relevant content. + 2. **If get_document returns None**: Use `await list_documents()` to see actual document titles, or `await search()` to find relevant content. 3. **Iterative Refinement**: Run code, examine results, adjust your approach based on what you find. - 4. **Use print() Liberally**: The REPL captures stdout - print intermediate results to see what you're working with. - 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and use collections. - 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use llm(). - 7. **Cite Your Sources**: Track which documents/chunks informed your answer for citation. - - ## DoclingDocument API - - When you call `get_docling_document(id_or_title)`, you get a DoclingDocument object for structured document analysis. - - ### Properties - - `doc.texts` - List of all text items (paragraphs, headings, etc.) - - `doc.tables` - List of all tables - - `doc.pictures` - List of all pictures/figures - - `doc.name` - Document name - - ### Methods - - `doc.iterate_items(with_groups=False)` - Iterate all items with hierarchy level - Returns tuples of (item, level) where level is nesting depth - - `doc.export_to_markdown()` - Export entire document as markdown string - - ### Text Item Properties - - `item.text` - The text content - - `item.label` - Type: title, paragraph, section_header, list_item, etc. (lowercase enum values) - - `item.prov` - Provenance (page numbers, bounding boxes) - - ### Table Access - - `table.data.num_rows`, `table.data.num_cols` - Dimensions - - `table.data.table_cells` - List of TableCell objects - - `cell.text`, `cell.start_row_offset_idx`, `cell.start_col_offset_idx` - - ### Example Usage - ```python - doc = get_docling_document("My Document") - - # Get all headings - headings = [t.text for t in doc.texts if "header" in str(t.label)] - - # Iterate with structure - for item, level in doc.iterate_items(): - print(" " * level + item.text[:50]) - - # Extract table data - for table in doc.tables: - for cell in table.data.table_cells: - print(f"Row {cell.start_row_offset_idx}, Col {cell.start_col_offset_idx}: {cell.text}") - ``` + 4. **Use print() Liberally**: The sandbox captures stdout - print intermediate results to see what you're working with. + 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and data structures. + 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use `await llm()`. + 7. **Cite Your Sources**: Use get_chunk() to retrieve chunk metadata for citations. Track which documents/chunks informed your answer. ## Example Patterns ### Counting documents matching a condition ```python - docs = list_documents(limit=100) + docs = await list_documents(limit=100) count = 0 for doc in docs: - content = get_document(doc['id']) + content = await get_document(doc['id']) if content and 'keyword' in content.lower(): count += 1 print(f"Found in: {doc['title']}") print(f"Total: {count}") ``` - ### Aggregating data across documents + ### Extracting data with llm() ```python - import re numbers = [] - results = search("financial data", limit=20) + results = await search("financial data", limit=20) for r in results: - matches = re.findall(r'\$([\d,]+)', r['content']) - for m in matches: - numbers.append(int(m.replace(',', ''))) - print(f"Average: ${sum(numbers)/len(numbers):,.2f}") + extracted = await llm(f"Extract all dollar amounts from this text as a comma-separated list of numbers (no $ signs): {r['content']}") + for part in extracted.split(','): + part = part.strip().replace(',', '') + if part.isdigit(): + numbers.append(int(part)) + if numbers: + print(f"Average: {sum(numbers) / len(numbers)}") + ``` + + ### Using search results with get_chunk for citations + ```python + results = await search("safety requirements", limit=5) + for r in results: + chunk = await get_chunk(r['chunk_id']) + print(f"From '{chunk['document_title']}', page {chunk['page_numbers']}: {chunk['content'][:100]}") ``` ### Using llm() for classification ```python - # Get document content - content = get_document("Q1 Report") - # Use llm() to classify sentiment - sentiment = llm(f"Classify the sentiment as positive, negative, or mixed: {content}") + content = await get_document("Q1 Report") + sentiment = await llm(f"Classify the sentiment as positive, negative, or mixed: {content}") print(sentiment) ``` @@ -260,11 +232,10 @@ interactions: tools: - function: description: |- - Execute Python code in a Docker-sandboxed environment. + Execute Python code in a sandboxed interpreter. The code has access to haiku.rag functions (search, list_documents, - get_document, get_docling_document, llm) and any Python standard - library module. + get_document, get_chunk, llm). Use print() to output results. @@ -304,7 +275,7 @@ interactions: response: headers: content-length: - - '622' + - '610' content-type: - application/json parsed_body: @@ -313,25 +284,25 @@ interactions: index: 0 message: content: '' - reasoning: Need to inspect documents variable. + reasoning: Need check documents variable? role: assistant tool_calls: - function: - arguments: '{"code":"if ''documents'' in dir():\n for doc in documents:\n print(doc[''title''], len(doc[''content'']))\nelse:\n print(''No - preloaded documents'')"}' + arguments: '{"code":"if ''documents'' in dir():\n for doc in documents:\n print(doc[''title''], len(doc[''content'']))\nelse:\n print(''no + documents'')\n"}' name: execute_code - id: call_dk6bzypo + id: call_2ao52bvz index: 0 type: function - created: 1770373463 - id: chatcmpl-869 + created: 1771924591 + id: chatcmpl-850 model: gpt-oss object: chat.completion system_fingerprint: fp_ollama usage: - completion_tokens: 69 - prompt_tokens: 1759 - total_tokens: 1828 + completion_tokens: 68 + prompt_tokens: 1635 + total_tokens: 1703 status: code: 200 message: OK @@ -344,7 +315,7 @@ interactions: connection: - keep-alive content-length: - - '8495' + - '8037' content-type: - application/json host: @@ -357,32 +328,33 @@ interactions: IMPORTANT: You MUST use the `execute_code` tool to run Python code. The functions described below are ONLY available inside the execute_code tool - you cannot access them any other way. Always execute code to answer questions; do not just describe what code would do. - CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just use them directly: - - search("query") ✓ CORRECT + CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just call them with `await`: + - results = await search("query") ✓ CORRECT - from haiku.rag import search ✗ WRONG - will fail + - results = search("query") ✗ WRONG - must use await - You have access to a sandboxed Python environment with these haiku.rag functions (use them directly, no imports needed): + You have access to a sandboxed Python interpreter with these haiku.rag functions (use them directly with `await`, no imports needed): ## Available Functions - ### search(query, limit=10) -> list[dict] + ### await search(query, limit=10) -> list[dict] Search the knowledge base using hybrid search (vector + full-text). Returns list of dicts with keys: chunk_id, content, document_id, document_title, document_uri, score, page_numbers, headings - ### list_documents(limit=10, offset=0) -> list[dict] + ### await list_documents(limit=10, offset=0) -> list[dict] List available documents in the knowledge base. Returns list of dicts with keys: id, title, uri, created_at - ### get_document(id_or_title) -> str | None + ### await get_document(id_or_title) -> str | None Get the full text content of a document by ID, title, or URI. Returns the document content as a string, or None if not found. - ### get_docling_document(id_or_title) -> DoclingDocument | None - Get the structured DoclingDocument object for advanced analysis. - Returns a DoclingDocument object, or None if not found. - See "DoclingDocument API" section below for how to use it. + ### await get_chunk(chunk_id) -> dict | None + Get a specific chunk by its ID (from search results). + Returns dict with keys: chunk_id, content, document_id, document_title, headings, page_numbers, labels + Use this to retrieve full chunk details and metadata for citation. - ### llm(prompt) -> str + ### await llm(prompt) -> str Call an LLM directly with the given prompt. Returns the response as a string. Use this for classification, summarization, extraction, or any task where you already have the content and just need LLM reasoning. @@ -397,93 +369,64 @@ interactions: ``` Check if it exists with: `if 'documents' in dir(): ...` - ## Standard Library Modules - You can import any Python standard library module. + ## Available Python Features + + The interpreter supports: variables, arithmetic, strings, f-strings, lists, dicts, tuples, sets, loops, conditionals, comprehensions, functions, async/await, `map()`, `sorted()`/`.sort(key=...)`, try/except, and the `json` module. + + Not supported: imports (other than `json`), class definitions, generators/yield, match statements, decorators, `with` statements. + + For pattern matching or text extraction, use string methods (`str.split`, `str.find`, `str.startswith`, `in` operator) or the `llm()` function. ## Strategy Guide 1. **Explore First**: Start by listing documents or searching to understand what's available. Document names may differ from filenames (e.g., "tbmed593.pdf" might be stored as "TB MED 593" or similar). - 2. **If get_document returns None**: Use `list_documents()` to see actual document titles, or `search()` to find relevant content. + 2. **If get_document returns None**: Use `await list_documents()` to see actual document titles, or `await search()` to find relevant content. 3. **Iterative Refinement**: Run code, examine results, adjust your approach based on what you find. - 4. **Use print() Liberally**: The REPL captures stdout - print intermediate results to see what you're working with. - 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and use collections. - 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use llm(). - 7. **Cite Your Sources**: Track which documents/chunks informed your answer for citation. - - ## DoclingDocument API - - When you call `get_docling_document(id_or_title)`, you get a DoclingDocument object for structured document analysis. - - ### Properties - - `doc.texts` - List of all text items (paragraphs, headings, etc.) - - `doc.tables` - List of all tables - - `doc.pictures` - List of all pictures/figures - - `doc.name` - Document name - - ### Methods - - `doc.iterate_items(with_groups=False)` - Iterate all items with hierarchy level - Returns tuples of (item, level) where level is nesting depth - - `doc.export_to_markdown()` - Export entire document as markdown string - - ### Text Item Properties - - `item.text` - The text content - - `item.label` - Type: title, paragraph, section_header, list_item, etc. (lowercase enum values) - - `item.prov` - Provenance (page numbers, bounding boxes) - - ### Table Access - - `table.data.num_rows`, `table.data.num_cols` - Dimensions - - `table.data.table_cells` - List of TableCell objects - - `cell.text`, `cell.start_row_offset_idx`, `cell.start_col_offset_idx` - - ### Example Usage - ```python - doc = get_docling_document("My Document") - - # Get all headings - headings = [t.text for t in doc.texts if "header" in str(t.label)] - - # Iterate with structure - for item, level in doc.iterate_items(): - print(" " * level + item.text[:50]) - - # Extract table data - for table in doc.tables: - for cell in table.data.table_cells: - print(f"Row {cell.start_row_offset_idx}, Col {cell.start_col_offset_idx}: {cell.text}") - ``` + 4. **Use print() Liberally**: The sandbox captures stdout - print intermediate results to see what you're working with. + 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and data structures. + 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use `await llm()`. + 7. **Cite Your Sources**: Use get_chunk() to retrieve chunk metadata for citations. Track which documents/chunks informed your answer. ## Example Patterns ### Counting documents matching a condition ```python - docs = list_documents(limit=100) + docs = await list_documents(limit=100) count = 0 for doc in docs: - content = get_document(doc['id']) + content = await get_document(doc['id']) if content and 'keyword' in content.lower(): count += 1 print(f"Found in: {doc['title']}") print(f"Total: {count}") ``` - ### Aggregating data across documents + ### Extracting data with llm() ```python - import re numbers = [] - results = search("financial data", limit=20) + results = await search("financial data", limit=20) for r in results: - matches = re.findall(r'\$([\d,]+)', r['content']) - for m in matches: - numbers.append(int(m.replace(',', ''))) - print(f"Average: ${sum(numbers)/len(numbers):,.2f}") + extracted = await llm(f"Extract all dollar amounts from this text as a comma-separated list of numbers (no $ signs): {r['content']}") + for part in extracted.split(','): + part = part.strip().replace(',', '') + if part.isdigit(): + numbers.append(int(part)) + if numbers: + print(f"Average: {sum(numbers) / len(numbers)}") + ``` + + ### Using search results with get_chunk for citations + ```python + results = await search("safety requirements", limit=5) + for r in results: + chunk = await get_chunk(r['chunk_id']) + print(f"From '{chunk['document_title']}', page {chunk['page_numbers']}: {chunk['content'][:100]}") ``` ### Using llm() for classification ```python - # Get document content - content = get_document("Q1 Report") - # Use llm() to classify sentiment - sentiment = llm(f"Classify the sentiment as positive, negative, or mixed: {content}") + content = await get_document("Q1 Report") + sentiment = await llm(f"Classify the sentiment as positive, negative, or mixed: {content}") print(sentiment) ``` @@ -509,22 +452,20 @@ interactions: role: system - content: Using the pre-loaded documents variable, tell me when was the company founded and what is their mission? role: user - - content: |- - - Need to inspect documents variable. - + - content: null + reasoning: Need check documents variable? role: assistant tool_calls: - function: - arguments: '{"code":"if ''documents'' in dir():\n for doc in documents:\n print(doc[''title''], len(doc[''content'']))\nelse:\n print(''No - preloaded documents'')"}' + arguments: '{"code":"if ''documents'' in dir():\n for doc in documents:\n print(doc[''title''], len(doc[''content'']))\nelse:\n print(''no + documents'')\n"}' name: execute_code - id: call_dk6bzypo + id: call_2ao52bvz type: function - - content: '{"code":"if ''documents'' in dir():\n for doc in documents:\n print(doc[''title''], len(doc[''content'']))\nelse:\n print(''No - preloaded documents'')","stdout":"No preloaded documents\n","stderr":"","success":true}' + - content: '{"code":"if ''documents'' in dir():\n for doc in documents:\n print(doc[''title''], len(doc[''content'']))\nelse:\n print(''no + documents'')\n","stdout":"","stderr":"NameError: name ''dir'' is not defined","success":false}' role: tool - tool_call_id: call_dk6bzypo + tool_call_id: call_2ao52bvz model: gpt-oss reasoning_effort: low stream: false @@ -532,11 +473,10 @@ interactions: tools: - function: description: |- - Execute Python code in a Docker-sandboxed environment. + Execute Python code in a sandboxed interpreter. The code has access to haiku.rag functions (search, list_documents, - get_document, get_docling_document, llm) and any Python standard - library module. + get_document, get_chunk, llm). Use print() to output results. @@ -576,7 +516,7 @@ interactions: response: headers: content-length: - - '661' + - '524' content-type: - application/json parsed_body: @@ -585,25 +525,24 @@ interactions: index: 0 message: content: '' - reasoning: No preloaded docs. Need to search. + reasoning: The environment may not provide dir(); use globals(). Check. role: assistant tool_calls: - function: - arguments: '{"code":"results = search(''company founded mission'', limit=10)\nfor r in results:\n print(''Title:'', - r[''document_title''], ''Score:'', r[''score''])\n print(r[''content''][:200])\n print(''---'')"}' + arguments: '{"code":"print(globals().keys())\n"}' name: execute_code - id: call_zf1l4x9e + id: call_rjj1hq9z index: 0 type: function - created: 1770373465 - id: chatcmpl-682 + created: 1771924592 + id: chatcmpl-798 model: gpt-oss object: chat.completion system_fingerprint: fp_ollama usage: - completion_tokens: 86 - prompt_tokens: 1903 - total_tokens: 1989 + completion_tokens: 43 + prompt_tokens: 1777 + total_tokens: 1820 status: code: 200 message: OK @@ -616,7 +555,7 @@ interactions: connection: - keep-alive content-length: - - '9434' + - '8486' content-type: - application/json host: @@ -629,32 +568,33 @@ interactions: IMPORTANT: You MUST use the `execute_code` tool to run Python code. The functions described below are ONLY available inside the execute_code tool - you cannot access them any other way. Always execute code to answer questions; do not just describe what code would do. - CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just use them directly: - - search("query") ✓ CORRECT + CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just call them with `await`: + - results = await search("query") ✓ CORRECT - from haiku.rag import search ✗ WRONG - will fail + - results = search("query") ✗ WRONG - must use await - You have access to a sandboxed Python environment with these haiku.rag functions (use them directly, no imports needed): + You have access to a sandboxed Python interpreter with these haiku.rag functions (use them directly with `await`, no imports needed): ## Available Functions - ### search(query, limit=10) -> list[dict] + ### await search(query, limit=10) -> list[dict] Search the knowledge base using hybrid search (vector + full-text). Returns list of dicts with keys: chunk_id, content, document_id, document_title, document_uri, score, page_numbers, headings - ### list_documents(limit=10, offset=0) -> list[dict] + ### await list_documents(limit=10, offset=0) -> list[dict] List available documents in the knowledge base. Returns list of dicts with keys: id, title, uri, created_at - ### get_document(id_or_title) -> str | None + ### await get_document(id_or_title) -> str | None Get the full text content of a document by ID, title, or URI. Returns the document content as a string, or None if not found. - ### get_docling_document(id_or_title) -> DoclingDocument | None - Get the structured DoclingDocument object for advanced analysis. - Returns a DoclingDocument object, or None if not found. - See "DoclingDocument API" section below for how to use it. + ### await get_chunk(chunk_id) -> dict | None + Get a specific chunk by its ID (from search results). + Returns dict with keys: chunk_id, content, document_id, document_title, headings, page_numbers, labels + Use this to retrieve full chunk details and metadata for citation. - ### llm(prompt) -> str + ### await llm(prompt) -> str Call an LLM directly with the given prompt. Returns the response as a string. Use this for classification, summarization, extraction, or any task where you already have the content and just need LLM reasoning. @@ -669,93 +609,64 @@ interactions: ``` Check if it exists with: `if 'documents' in dir(): ...` - ## Standard Library Modules - You can import any Python standard library module. + ## Available Python Features + + The interpreter supports: variables, arithmetic, strings, f-strings, lists, dicts, tuples, sets, loops, conditionals, comprehensions, functions, async/await, `map()`, `sorted()`/`.sort(key=...)`, try/except, and the `json` module. + + Not supported: imports (other than `json`), class definitions, generators/yield, match statements, decorators, `with` statements. + + For pattern matching or text extraction, use string methods (`str.split`, `str.find`, `str.startswith`, `in` operator) or the `llm()` function. ## Strategy Guide 1. **Explore First**: Start by listing documents or searching to understand what's available. Document names may differ from filenames (e.g., "tbmed593.pdf" might be stored as "TB MED 593" or similar). - 2. **If get_document returns None**: Use `list_documents()` to see actual document titles, or `search()` to find relevant content. + 2. **If get_document returns None**: Use `await list_documents()` to see actual document titles, or `await search()` to find relevant content. 3. **Iterative Refinement**: Run code, examine results, adjust your approach based on what you find. - 4. **Use print() Liberally**: The REPL captures stdout - print intermediate results to see what you're working with. - 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and use collections. - 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use llm(). - 7. **Cite Your Sources**: Track which documents/chunks informed your answer for citation. - - ## DoclingDocument API - - When you call `get_docling_document(id_or_title)`, you get a DoclingDocument object for structured document analysis. - - ### Properties - - `doc.texts` - List of all text items (paragraphs, headings, etc.) - - `doc.tables` - List of all tables - - `doc.pictures` - List of all pictures/figures - - `doc.name` - Document name - - ### Methods - - `doc.iterate_items(with_groups=False)` - Iterate all items with hierarchy level - Returns tuples of (item, level) where level is nesting depth - - `doc.export_to_markdown()` - Export entire document as markdown string - - ### Text Item Properties - - `item.text` - The text content - - `item.label` - Type: title, paragraph, section_header, list_item, etc. (lowercase enum values) - - `item.prov` - Provenance (page numbers, bounding boxes) - - ### Table Access - - `table.data.num_rows`, `table.data.num_cols` - Dimensions - - `table.data.table_cells` - List of TableCell objects - - `cell.text`, `cell.start_row_offset_idx`, `cell.start_col_offset_idx` - - ### Example Usage - ```python - doc = get_docling_document("My Document") - - # Get all headings - headings = [t.text for t in doc.texts if "header" in str(t.label)] - - # Iterate with structure - for item, level in doc.iterate_items(): - print(" " * level + item.text[:50]) - - # Extract table data - for table in doc.tables: - for cell in table.data.table_cells: - print(f"Row {cell.start_row_offset_idx}, Col {cell.start_col_offset_idx}: {cell.text}") - ``` + 4. **Use print() Liberally**: The sandbox captures stdout - print intermediate results to see what you're working with. + 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and data structures. + 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use `await llm()`. + 7. **Cite Your Sources**: Use get_chunk() to retrieve chunk metadata for citations. Track which documents/chunks informed your answer. ## Example Patterns ### Counting documents matching a condition ```python - docs = list_documents(limit=100) + docs = await list_documents(limit=100) count = 0 for doc in docs: - content = get_document(doc['id']) + content = await get_document(doc['id']) if content and 'keyword' in content.lower(): count += 1 print(f"Found in: {doc['title']}") print(f"Total: {count}") ``` - ### Aggregating data across documents + ### Extracting data with llm() ```python - import re numbers = [] - results = search("financial data", limit=20) + results = await search("financial data", limit=20) for r in results: - matches = re.findall(r'\$([\d,]+)', r['content']) - for m in matches: - numbers.append(int(m.replace(',', ''))) - print(f"Average: ${sum(numbers)/len(numbers):,.2f}") + extracted = await llm(f"Extract all dollar amounts from this text as a comma-separated list of numbers (no $ signs): {r['content']}") + for part in extracted.split(','): + part = part.strip().replace(',', '') + if part.isdigit(): + numbers.append(int(part)) + if numbers: + print(f"Average: {sum(numbers) / len(numbers)}") + ``` + + ### Using search results with get_chunk for citations + ```python + results = await search("safety requirements", limit=5) + for r in results: + chunk = await get_chunk(r['chunk_id']) + print(f"From '{chunk['document_title']}', page {chunk['page_numbers']}: {chunk['content'][:100]}") ``` ### Using llm() for classification ```python - # Get document content - content = get_document("Q1 Report") - # Use llm() to classify sentiment - sentiment = llm(f"Classify the sentiment as positive, negative, or mixed: {content}") + content = await get_document("Q1 Report") + sentiment = await llm(f"Classify the sentiment as positive, negative, or mixed: {content}") print(sentiment) ``` @@ -781,40 +692,32 @@ interactions: role: system - content: Using the pre-loaded documents variable, tell me when was the company founded and what is their mission? role: user - - content: |- - - Need to inspect documents variable. - + - content: null + reasoning: Need check documents variable? role: assistant tool_calls: - function: - arguments: '{"code":"if ''documents'' in dir():\n for doc in documents:\n print(doc[''title''], len(doc[''content'']))\nelse:\n print(''No - preloaded documents'')"}' + arguments: '{"code":"if ''documents'' in dir():\n for doc in documents:\n print(doc[''title''], len(doc[''content'']))\nelse:\n print(''no + documents'')\n"}' name: execute_code - id: call_dk6bzypo + id: call_2ao52bvz type: function - - content: '{"code":"if ''documents'' in dir():\n for doc in documents:\n print(doc[''title''], len(doc[''content'']))\nelse:\n print(''No - preloaded documents'')","stdout":"No preloaded documents\n","stderr":"","success":true}' + - content: '{"code":"if ''documents'' in dir():\n for doc in documents:\n print(doc[''title''], len(doc[''content'']))\nelse:\n print(''no + documents'')\n","stdout":"","stderr":"NameError: name ''dir'' is not defined","success":false}' role: tool - tool_call_id: call_dk6bzypo - - content: |- - - No preloaded docs. Need to search. - + tool_call_id: call_2ao52bvz + - content: null + reasoning: The environment may not provide dir(); use globals(). Check. role: assistant tool_calls: - function: - arguments: '{"code":"results = search(''company founded mission'', limit=10)\nfor r in results:\n print(''Title:'', - r[''document_title''], ''Score:'', r[''score''])\n print(r[''content''][:200])\n print(''---'')"}' + arguments: '{"code":"print(globals().keys())\n"}' name: execute_code - id: call_zf1l4x9e + id: call_rjj1hq9z type: function - - content: '{"code":"results = search(''company founded mission'', limit=10)\nfor r in results:\n print(''Title:'', - r[''document_title''], ''Score:'', r[''score''])\n print(r[''content''][:200])\n print(''---'')","stdout":"Title: - Mission Statement Score: 0.032522473484277725\nOur mission is to make technology accessible to everyone.\n---\nTitle: - Company History Score: 0.032522473484277725\nThe company was founded in 1985 by Jane Smith.\n---\n","stderr":"","success":true}' + - content: '{"code":"print(globals().keys())\n","stdout":"","stderr":"NameError: name ''globals'' is not defined","success":false}' role: tool - tool_call_id: call_zf1l4x9e + tool_call_id: call_rjj1hq9z model: gpt-oss reasoning_effort: low stream: false @@ -822,11 +725,10 @@ interactions: tools: - function: description: |- - Execute Python code in a Docker-sandboxed environment. + Execute Python code in a sandboxed interpreter. The code has access to haiku.rag functions (search, list_documents, - get_document, get_docling_document, llm) and any Python standard - library module. + get_document, get_chunk, llm). Use print() to output results. @@ -866,7 +768,1544 @@ interactions: response: headers: content-length: - - '1256' + - '581' + content-type: + - application/json + parsed_body: + choices: + - finish_reason: tool_calls + index: 0 + message: + content: '' + reasoning: Probably no python environment? Might need simple print? Might fail. Probably cannot see documents. Could + search. + role: assistant + tool_calls: + - function: + arguments: '{"query":"founded mission company","limit":5}' + name: search + id: call_ee8bn4w7 + index: 0 + type: function + created: 1771924593 + id: chatcmpl-566 + model: gpt-oss + object: chat.completion + system_fingerprint: fp_ollama + usage: + completion_tokens: 51 + prompt_tokens: 1862 + total_tokens: 1913 + status: + code: 200 + message: OK +- request: + headers: + accept: + - application/json + accept-encoding: + - gzip, deflate, zstd + connection: + - keep-alive + content-length: + - '8970' + content-type: + - application/json + host: + - localhost:11434 + method: POST + parsed_body: + messages: + - content: |- + You are a Recursive Language Model (RLM) agent that solves complex research questions by writing and executing Python code. + + IMPORTANT: You MUST use the `execute_code` tool to run Python code. The functions described below are ONLY available inside the execute_code tool - you cannot access them any other way. Always execute code to answer questions; do not just describe what code would do. + + CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just call them with `await`: + - results = await search("query") ✓ CORRECT + - from haiku.rag import search ✗ WRONG - will fail + - results = search("query") ✗ WRONG - must use await + + You have access to a sandboxed Python interpreter with these haiku.rag functions (use them directly with `await`, no imports needed): + + ## Available Functions + + ### await search(query, limit=10) -> list[dict] + Search the knowledge base using hybrid search (vector + full-text). + Returns list of dicts with keys: chunk_id, content, document_id, document_title, document_uri, score, page_numbers, headings + + ### await list_documents(limit=10, offset=0) -> list[dict] + List available documents in the knowledge base. + Returns list of dicts with keys: id, title, uri, created_at + + ### await get_document(id_or_title) -> str | None + Get the full text content of a document by ID, title, or URI. + Returns the document content as a string, or None if not found. + + ### await get_chunk(chunk_id) -> dict | None + Get a specific chunk by its ID (from search results). + Returns dict with keys: chunk_id, content, document_id, document_title, headings, page_numbers, labels + Use this to retrieve full chunk details and metadata for citation. + + ### await llm(prompt) -> str + Call an LLM directly with the given prompt. Returns the response as a string. + Use this for classification, summarization, extraction, or any task where you + already have the content and just need LLM reasoning. + + ## Pre-loaded Documents Variable + + If documents were pre-loaded for this session, a `documents` variable is available: + ```python + # documents is a list of dicts with keys: id, title, uri, content + for doc in documents: + print(doc['title'], len(doc['content'])) + ``` + Check if it exists with: `if 'documents' in dir(): ...` + + ## Available Python Features + + The interpreter supports: variables, arithmetic, strings, f-strings, lists, dicts, tuples, sets, loops, conditionals, comprehensions, functions, async/await, `map()`, `sorted()`/`.sort(key=...)`, try/except, and the `json` module. + + Not supported: imports (other than `json`), class definitions, generators/yield, match statements, decorators, `with` statements. + + For pattern matching or text extraction, use string methods (`str.split`, `str.find`, `str.startswith`, `in` operator) or the `llm()` function. + + ## Strategy Guide + + 1. **Explore First**: Start by listing documents or searching to understand what's available. Document names may differ from filenames (e.g., "tbmed593.pdf" might be stored as "TB MED 593" or similar). + 2. **If get_document returns None**: Use `await list_documents()` to see actual document titles, or `await search()` to find relevant content. + 3. **Iterative Refinement**: Run code, examine results, adjust your approach based on what you find. + 4. **Use print() Liberally**: The sandbox captures stdout - print intermediate results to see what you're working with. + 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and data structures. + 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use `await llm()`. + 7. **Cite Your Sources**: Use get_chunk() to retrieve chunk metadata for citations. Track which documents/chunks informed your answer. + + ## Example Patterns + + ### Counting documents matching a condition + ```python + docs = await list_documents(limit=100) + count = 0 + for doc in docs: + content = await get_document(doc['id']) + if content and 'keyword' in content.lower(): + count += 1 + print(f"Found in: {doc['title']}") + print(f"Total: {count}") + ``` + + ### Extracting data with llm() + ```python + numbers = [] + results = await search("financial data", limit=20) + for r in results: + extracted = await llm(f"Extract all dollar amounts from this text as a comma-separated list of numbers (no $ signs): {r['content']}") + for part in extracted.split(','): + part = part.strip().replace(',', '') + if part.isdigit(): + numbers.append(int(part)) + if numbers: + print(f"Average: {sum(numbers) / len(numbers)}") + ``` + + ### Using search results with get_chunk for citations + ```python + results = await search("safety requirements", limit=5) + for r in results: + chunk = await get_chunk(r['chunk_id']) + print(f"From '{chunk['document_title']}', page {chunk['page_numbers']}: {chunk['content'][:100]}") + ``` + + ### Using llm() for classification + ```python + content = await get_document("Q1 Report") + sentiment = await llm(f"Classify the sentiment as positive, negative, or mixed: {content}") + print(sentiment) + ``` + + ## Workflow + + 1. **ALWAYS start by using execute_code** to explore the knowledge base + 2. Run multiple code blocks as needed to gather information + 3. After collecting data, provide your final answer + + ## Output Format + + CRITICAL: Your final response MUST be valid JSON matching this exact schema: + ```json + {"answer": "Your complete answer here as a string", "program": "Your final consolidated program here as a string"} + ``` + + - `answer`: A clear answer to the user's question with key findings and references to specific documents/chunks. + - `program`: A single, self-contained Python program that produces the answer. Consolidate your exploratory code executions into one clean script. + + Do NOT return arbitrary JSON structures. Always use the exact format: {"answer": "...", "program": "..."} + + CRITICAL: You MUST call execute_code at least once before providing your answer. Never give up without trying to execute code first. + role: system + - content: Using the pre-loaded documents variable, tell me when was the company founded and what is their mission? + role: user + - content: null + reasoning: Need check documents variable? + role: assistant + tool_calls: + - function: + arguments: '{"code":"if ''documents'' in dir():\n for doc in documents:\n print(doc[''title''], len(doc[''content'']))\nelse:\n print(''no + documents'')\n"}' + name: execute_code + id: call_2ao52bvz + type: function + - content: '{"code":"if ''documents'' in dir():\n for doc in documents:\n print(doc[''title''], len(doc[''content'']))\nelse:\n print(''no + documents'')\n","stdout":"","stderr":"NameError: name ''dir'' is not defined","success":false}' + role: tool + tool_call_id: call_2ao52bvz + - content: null + reasoning: The environment may not provide dir(); use globals(). Check. + role: assistant + tool_calls: + - function: + arguments: '{"code":"print(globals().keys())\n"}' + name: execute_code + id: call_rjj1hq9z + type: function + - content: '{"code":"print(globals().keys())\n","stdout":"","stderr":"NameError: name ''globals'' is not defined","success":false}' + role: tool + tool_call_id: call_rjj1hq9z + - content: null + reasoning: Probably no python environment? Might need simple print? Might fail. Probably cannot see documents. Could + search. + role: assistant + tool_calls: + - function: + arguments: '{"query":"founded mission company","limit":5}' + name: search + id: call_ee8bn4w7 + type: function + - content: |- + Unknown tool name: 'search'. Available tools: 'final_result', 'execute_code' + + Fix the errors and try again. + role: tool + tool_call_id: call_ee8bn4w7 + model: gpt-oss + reasoning_effort: low + stream: false + tool_choice: auto + tools: + - function: + description: |- + Execute Python code in a sandboxed interpreter. + + The code has access to haiku.rag functions (search, list_documents, + get_document, get_chunk, llm). + + Use print() to output results. + + Structured result with success status, stdout, and stderr. + + name: execute_code + parameters: + additionalProperties: false + properties: + code: + description: Python code to execute. + type: string + required: + - code + type: object + strict: true + type: function + - function: + description: Result from RLM agent execution. + name: final_result + parameters: + additionalProperties: false + properties: + answer: + description: The answer to the user's question + type: string + program: + description: The final consolidated program + type: string + required: + - answer + - program + type: object + strict: true + type: function + uri: http://localhost:11434/v1/chat/completions + response: + headers: + content-length: + - '550' + content-type: + - application/json + parsed_body: + choices: + - finish_reason: tool_calls + index: 0 + message: + content: '' + reasoning: The tool is execute_code. Use that. + role: assistant + tool_calls: + - function: + arguments: '{"code":"results = await search(''founded mission company'', limit=5)\nprint(results)\n"}' + name: execute_code + id: call_o1c87rz5 + index: 0 + type: function + created: 1771924595 + id: chatcmpl-37 + model: gpt-oss + object: chat.completion + system_fingerprint: fp_ollama + usage: + completion_tokens: 51 + prompt_tokens: 1948 + total_tokens: 1999 + status: + code: 200 + message: OK +- request: + headers: + accept: + - application/json + accept-encoding: + - gzip, deflate, zstd + connection: + - keep-alive + content-length: + - '93' + content-type: + - application/json + host: + - localhost:11434 + method: POST + parsed_body: + encoding_format: base64 + input: + - founded mission company + model: qwen3-embedding:4b + uri: http://localhost:11434/v1/embeddings + response: + headers: + content-type: + - application/json + transfer-encoding: + - chunked + parsed_body: + data: + - embedding: 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 + index: 0 + object: embedding + model: qwen3-embedding:4b + object: list + usage: + prompt_tokens: 5 + total_tokens: 5 + status: + code: 200 + message: OK +- request: + headers: + accept: + - application/json + accept-encoding: + - gzip, deflate, zstd + connection: + - keep-alive + content-length: + - '10069' + content-type: + - application/json + host: + - localhost:11434 + method: POST + parsed_body: + messages: + - content: |- + You are a Recursive Language Model (RLM) agent that solves complex research questions by writing and executing Python code. + + IMPORTANT: You MUST use the `execute_code` tool to run Python code. The functions described below are ONLY available inside the execute_code tool - you cannot access them any other way. Always execute code to answer questions; do not just describe what code would do. + + CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just call them with `await`: + - results = await search("query") ✓ CORRECT + - from haiku.rag import search ✗ WRONG - will fail + - results = search("query") ✗ WRONG - must use await + + You have access to a sandboxed Python interpreter with these haiku.rag functions (use them directly with `await`, no imports needed): + + ## Available Functions + + ### await search(query, limit=10) -> list[dict] + Search the knowledge base using hybrid search (vector + full-text). + Returns list of dicts with keys: chunk_id, content, document_id, document_title, document_uri, score, page_numbers, headings + + ### await list_documents(limit=10, offset=0) -> list[dict] + List available documents in the knowledge base. + Returns list of dicts with keys: id, title, uri, created_at + + ### await get_document(id_or_title) -> str | None + Get the full text content of a document by ID, title, or URI. + Returns the document content as a string, or None if not found. + + ### await get_chunk(chunk_id) -> dict | None + Get a specific chunk by its ID (from search results). + Returns dict with keys: chunk_id, content, document_id, document_title, headings, page_numbers, labels + Use this to retrieve full chunk details and metadata for citation. + + ### await llm(prompt) -> str + Call an LLM directly with the given prompt. Returns the response as a string. + Use this for classification, summarization, extraction, or any task where you + already have the content and just need LLM reasoning. + + ## Pre-loaded Documents Variable + + If documents were pre-loaded for this session, a `documents` variable is available: + ```python + # documents is a list of dicts with keys: id, title, uri, content + for doc in documents: + print(doc['title'], len(doc['content'])) + ``` + Check if it exists with: `if 'documents' in dir(): ...` + + ## Available Python Features + + The interpreter supports: variables, arithmetic, strings, f-strings, lists, dicts, tuples, sets, loops, conditionals, comprehensions, functions, async/await, `map()`, `sorted()`/`.sort(key=...)`, try/except, and the `json` module. + + Not supported: imports (other than `json`), class definitions, generators/yield, match statements, decorators, `with` statements. + + For pattern matching or text extraction, use string methods (`str.split`, `str.find`, `str.startswith`, `in` operator) or the `llm()` function. + + ## Strategy Guide + + 1. **Explore First**: Start by listing documents or searching to understand what's available. Document names may differ from filenames (e.g., "tbmed593.pdf" might be stored as "TB MED 593" or similar). + 2. **If get_document returns None**: Use `await list_documents()` to see actual document titles, or `await search()` to find relevant content. + 3. **Iterative Refinement**: Run code, examine results, adjust your approach based on what you find. + 4. **Use print() Liberally**: The sandbox captures stdout - print intermediate results to see what you're working with. + 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and data structures. + 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use `await llm()`. + 7. **Cite Your Sources**: Use get_chunk() to retrieve chunk metadata for citations. Track which documents/chunks informed your answer. + + ## Example Patterns + + ### Counting documents matching a condition + ```python + docs = await list_documents(limit=100) + count = 0 + for doc in docs: + content = await get_document(doc['id']) + if content and 'keyword' in content.lower(): + count += 1 + print(f"Found in: {doc['title']}") + print(f"Total: {count}") + ``` + + ### Extracting data with llm() + ```python + numbers = [] + results = await search("financial data", limit=20) + for r in results: + extracted = await llm(f"Extract all dollar amounts from this text as a comma-separated list of numbers (no $ signs): {r['content']}") + for part in extracted.split(','): + part = part.strip().replace(',', '') + if part.isdigit(): + numbers.append(int(part)) + if numbers: + print(f"Average: {sum(numbers) / len(numbers)}") + ``` + + ### Using search results with get_chunk for citations + ```python + results = await search("safety requirements", limit=5) + for r in results: + chunk = await get_chunk(r['chunk_id']) + print(f"From '{chunk['document_title']}', page {chunk['page_numbers']}: {chunk['content'][:100]}") + ``` + + ### Using llm() for classification + ```python + content = await get_document("Q1 Report") + sentiment = await llm(f"Classify the sentiment as positive, negative, or mixed: {content}") + print(sentiment) + ``` + + ## Workflow + + 1. **ALWAYS start by using execute_code** to explore the knowledge base + 2. Run multiple code blocks as needed to gather information + 3. After collecting data, provide your final answer + + ## Output Format + + CRITICAL: Your final response MUST be valid JSON matching this exact schema: + ```json + {"answer": "Your complete answer here as a string", "program": "Your final consolidated program here as a string"} + ``` + + - `answer`: A clear answer to the user's question with key findings and references to specific documents/chunks. + - `program`: A single, self-contained Python program that produces the answer. Consolidate your exploratory code executions into one clean script. + + Do NOT return arbitrary JSON structures. Always use the exact format: {"answer": "...", "program": "..."} + + CRITICAL: You MUST call execute_code at least once before providing your answer. Never give up without trying to execute code first. + role: system + - content: Using the pre-loaded documents variable, tell me when was the company founded and what is their mission? + role: user + - content: null + reasoning: Need check documents variable? + role: assistant + tool_calls: + - function: + arguments: '{"code":"if ''documents'' in dir():\n for doc in documents:\n print(doc[''title''], len(doc[''content'']))\nelse:\n print(''no + documents'')\n"}' + name: execute_code + id: call_2ao52bvz + type: function + - content: '{"code":"if ''documents'' in dir():\n for doc in documents:\n print(doc[''title''], len(doc[''content'']))\nelse:\n print(''no + documents'')\n","stdout":"","stderr":"NameError: name ''dir'' is not defined","success":false}' + role: tool + tool_call_id: call_2ao52bvz + - content: null + reasoning: The environment may not provide dir(); use globals(). Check. + role: assistant + tool_calls: + - function: + arguments: '{"code":"print(globals().keys())\n"}' + name: execute_code + id: call_rjj1hq9z + type: function + - content: '{"code":"print(globals().keys())\n","stdout":"","stderr":"NameError: name ''globals'' is not defined","success":false}' + role: tool + tool_call_id: call_rjj1hq9z + - content: null + reasoning: Probably no python environment? Might need simple print? Might fail. Probably cannot see documents. Could + search. + role: assistant + tool_calls: + - function: + arguments: '{"query":"founded mission company","limit":5}' + name: search + id: call_ee8bn4w7 + type: function + - content: |- + Unknown tool name: 'search'. Available tools: 'final_result', 'execute_code' + + Fix the errors and try again. + role: tool + tool_call_id: call_ee8bn4w7 + - content: null + reasoning: The tool is execute_code. Use that. + role: assistant + tool_calls: + - function: + arguments: '{"code":"results = await search(''founded mission company'', limit=5)\nprint(results)\n"}' + name: execute_code + id: call_o1c87rz5 + type: function + - content: '{"code":"results = await search(''founded mission company'', limit=5)\nprint(results)\n","stdout":"[{''chunk_id'': + ''d43b4445-8425-4434-bc36-e28e6f7e573e'', ''content'': ''Our mission is to make technology accessible to everyone.'', + ''document_id'': ''bb1e588e-8014-4607-bf9b-a077c16d1cdf'', ''document_title'': ''Mission Statement'', ''document_uri'': + None, ''score'': 0.032522473484277725, ''page_numbers'': [], ''headings'': None}, {''chunk_id'': ''6d291645-590e-498c-a22c-0883e7b1582b'', + ''content'': ''The company was founded in 1985 by Jane Smith.'', ''document_id'': ''a2e1e13c-e316-4198-a5a8-49353eaa7337'', + ''document_title'': ''Company History'', ''document_uri'': None, ''score'': 0.032522473484277725, ''page_numbers'': + [], ''headings'': None}]\n","stderr":"","success":true}' + role: tool + tool_call_id: call_o1c87rz5 + model: gpt-oss + reasoning_effort: low + stream: false + tool_choice: auto + tools: + - function: + description: |- + Execute Python code in a sandboxed interpreter. + + The code has access to haiku.rag functions (search, list_documents, + get_document, get_chunk, llm). + + Use print() to output results. + + Structured result with success status, stdout, and stderr. + + name: execute_code + parameters: + additionalProperties: false + properties: + code: + description: Python code to execute. + type: string + required: + - code + type: object + strict: true + type: function + - function: + description: Result from RLM agent execution. + name: final_result + parameters: + additionalProperties: false + properties: + answer: + description: The answer to the user's question + type: string + program: + description: The final consolidated program + type: string + required: + - answer + - program + type: object + strict: true + type: function + uri: http://localhost:11434/v1/chat/completions + response: + headers: + content-length: + - '509' + content-type: + - application/json + parsed_body: + choices: + - finish_reason: tool_calls + index: 0 + message: + content: '' + reasoning: Got info. Need citation get_chunk. + role: assistant + tool_calls: + - function: + arguments: '{"chunk_id":"6d291645-590e-498c-a22c-0883e7b1582b"}' + name: get_chunk + id: call_b0i0vdgg + index: 0 + type: function + created: 1771924596 + id: chatcmpl-840 + model: gpt-oss + object: chat.completion + system_fingerprint: fp_ollama + usage: + completion_tokens: 54 + prompt_tokens: 2265 + total_tokens: 2319 + status: + code: 200 + message: OK +- request: + headers: + accept: + - application/json + accept-encoding: + - gzip, deflate, zstd + connection: + - keep-alive + content-length: + - '10484' + content-type: + - application/json + host: + - localhost:11434 + method: POST + parsed_body: + messages: + - content: |- + You are a Recursive Language Model (RLM) agent that solves complex research questions by writing and executing Python code. + + IMPORTANT: You MUST use the `execute_code` tool to run Python code. The functions described below are ONLY available inside the execute_code tool - you cannot access them any other way. Always execute code to answer questions; do not just describe what code would do. + + CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just call them with `await`: + - results = await search("query") ✓ CORRECT + - from haiku.rag import search ✗ WRONG - will fail + - results = search("query") ✗ WRONG - must use await + + You have access to a sandboxed Python interpreter with these haiku.rag functions (use them directly with `await`, no imports needed): + + ## Available Functions + + ### await search(query, limit=10) -> list[dict] + Search the knowledge base using hybrid search (vector + full-text). + Returns list of dicts with keys: chunk_id, content, document_id, document_title, document_uri, score, page_numbers, headings + + ### await list_documents(limit=10, offset=0) -> list[dict] + List available documents in the knowledge base. + Returns list of dicts with keys: id, title, uri, created_at + + ### await get_document(id_or_title) -> str | None + Get the full text content of a document by ID, title, or URI. + Returns the document content as a string, or None if not found. + + ### await get_chunk(chunk_id) -> dict | None + Get a specific chunk by its ID (from search results). + Returns dict with keys: chunk_id, content, document_id, document_title, headings, page_numbers, labels + Use this to retrieve full chunk details and metadata for citation. + + ### await llm(prompt) -> str + Call an LLM directly with the given prompt. Returns the response as a string. + Use this for classification, summarization, extraction, or any task where you + already have the content and just need LLM reasoning. + + ## Pre-loaded Documents Variable + + If documents were pre-loaded for this session, a `documents` variable is available: + ```python + # documents is a list of dicts with keys: id, title, uri, content + for doc in documents: + print(doc['title'], len(doc['content'])) + ``` + Check if it exists with: `if 'documents' in dir(): ...` + + ## Available Python Features + + The interpreter supports: variables, arithmetic, strings, f-strings, lists, dicts, tuples, sets, loops, conditionals, comprehensions, functions, async/await, `map()`, `sorted()`/`.sort(key=...)`, try/except, and the `json` module. + + Not supported: imports (other than `json`), class definitions, generators/yield, match statements, decorators, `with` statements. + + For pattern matching or text extraction, use string methods (`str.split`, `str.find`, `str.startswith`, `in` operator) or the `llm()` function. + + ## Strategy Guide + + 1. **Explore First**: Start by listing documents or searching to understand what's available. Document names may differ from filenames (e.g., "tbmed593.pdf" might be stored as "TB MED 593" or similar). + 2. **If get_document returns None**: Use `await list_documents()` to see actual document titles, or `await search()` to find relevant content. + 3. **Iterative Refinement**: Run code, examine results, adjust your approach based on what you find. + 4. **Use print() Liberally**: The sandbox captures stdout - print intermediate results to see what you're working with. + 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and data structures. + 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use `await llm()`. + 7. **Cite Your Sources**: Use get_chunk() to retrieve chunk metadata for citations. Track which documents/chunks informed your answer. + + ## Example Patterns + + ### Counting documents matching a condition + ```python + docs = await list_documents(limit=100) + count = 0 + for doc in docs: + content = await get_document(doc['id']) + if content and 'keyword' in content.lower(): + count += 1 + print(f"Found in: {doc['title']}") + print(f"Total: {count}") + ``` + + ### Extracting data with llm() + ```python + numbers = [] + results = await search("financial data", limit=20) + for r in results: + extracted = await llm(f"Extract all dollar amounts from this text as a comma-separated list of numbers (no $ signs): {r['content']}") + for part in extracted.split(','): + part = part.strip().replace(',', '') + if part.isdigit(): + numbers.append(int(part)) + if numbers: + print(f"Average: {sum(numbers) / len(numbers)}") + ``` + + ### Using search results with get_chunk for citations + ```python + results = await search("safety requirements", limit=5) + for r in results: + chunk = await get_chunk(r['chunk_id']) + print(f"From '{chunk['document_title']}', page {chunk['page_numbers']}: {chunk['content'][:100]}") + ``` + + ### Using llm() for classification + ```python + content = await get_document("Q1 Report") + sentiment = await llm(f"Classify the sentiment as positive, negative, or mixed: {content}") + print(sentiment) + ``` + + ## Workflow + + 1. **ALWAYS start by using execute_code** to explore the knowledge base + 2. Run multiple code blocks as needed to gather information + 3. After collecting data, provide your final answer + + ## Output Format + + CRITICAL: Your final response MUST be valid JSON matching this exact schema: + ```json + {"answer": "Your complete answer here as a string", "program": "Your final consolidated program here as a string"} + ``` + + - `answer`: A clear answer to the user's question with key findings and references to specific documents/chunks. + - `program`: A single, self-contained Python program that produces the answer. Consolidate your exploratory code executions into one clean script. + + Do NOT return arbitrary JSON structures. Always use the exact format: {"answer": "...", "program": "..."} + + CRITICAL: You MUST call execute_code at least once before providing your answer. Never give up without trying to execute code first. + role: system + - content: Using the pre-loaded documents variable, tell me when was the company founded and what is their mission? + role: user + - content: null + reasoning: Need check documents variable? + role: assistant + tool_calls: + - function: + arguments: '{"code":"if ''documents'' in dir():\n for doc in documents:\n print(doc[''title''], len(doc[''content'']))\nelse:\n print(''no + documents'')\n"}' + name: execute_code + id: call_2ao52bvz + type: function + - content: '{"code":"if ''documents'' in dir():\n for doc in documents:\n print(doc[''title''], len(doc[''content'']))\nelse:\n print(''no + documents'')\n","stdout":"","stderr":"NameError: name ''dir'' is not defined","success":false}' + role: tool + tool_call_id: call_2ao52bvz + - content: null + reasoning: The environment may not provide dir(); use globals(). Check. + role: assistant + tool_calls: + - function: + arguments: '{"code":"print(globals().keys())\n"}' + name: execute_code + id: call_rjj1hq9z + type: function + - content: '{"code":"print(globals().keys())\n","stdout":"","stderr":"NameError: name ''globals'' is not defined","success":false}' + role: tool + tool_call_id: call_rjj1hq9z + - content: null + reasoning: Probably no python environment? Might need simple print? Might fail. Probably cannot see documents. Could + search. + role: assistant + tool_calls: + - function: + arguments: '{"query":"founded mission company","limit":5}' + name: search + id: call_ee8bn4w7 + type: function + - content: |- + Unknown tool name: 'search'. Available tools: 'final_result', 'execute_code' + + Fix the errors and try again. + role: tool + tool_call_id: call_ee8bn4w7 + - content: null + reasoning: The tool is execute_code. Use that. + role: assistant + tool_calls: + - function: + arguments: '{"code":"results = await search(''founded mission company'', limit=5)\nprint(results)\n"}' + name: execute_code + id: call_o1c87rz5 + type: function + - content: '{"code":"results = await search(''founded mission company'', limit=5)\nprint(results)\n","stdout":"[{''chunk_id'': + ''d43b4445-8425-4434-bc36-e28e6f7e573e'', ''content'': ''Our mission is to make technology accessible to everyone.'', + ''document_id'': ''bb1e588e-8014-4607-bf9b-a077c16d1cdf'', ''document_title'': ''Mission Statement'', ''document_uri'': + None, ''score'': 0.032522473484277725, ''page_numbers'': [], ''headings'': None}, {''chunk_id'': ''6d291645-590e-498c-a22c-0883e7b1582b'', + ''content'': ''The company was founded in 1985 by Jane Smith.'', ''document_id'': ''a2e1e13c-e316-4198-a5a8-49353eaa7337'', + ''document_title'': ''Company History'', ''document_uri'': None, ''score'': 0.032522473484277725, ''page_numbers'': + [], ''headings'': None}]\n","stderr":"","success":true}' + role: tool + tool_call_id: call_o1c87rz5 + - content: null + reasoning: Got info. Need citation get_chunk. + role: assistant + tool_calls: + - function: + arguments: '{"chunk_id":"6d291645-590e-498c-a22c-0883e7b1582b"}' + name: get_chunk + id: call_b0i0vdgg + type: function + - content: |- + Unknown tool name: 'get_chunk'. Available tools: 'final_result', 'execute_code' + + Fix the errors and try again. + role: tool + tool_call_id: call_b0i0vdgg + model: gpt-oss + reasoning_effort: low + stream: false + tool_choice: auto + tools: + - function: + description: |- + Execute Python code in a sandboxed interpreter. + + The code has access to haiku.rag functions (search, list_documents, + get_document, get_chunk, llm). + + Use print() to output results. + + Structured result with success status, stdout, and stderr. + + name: execute_code + parameters: + additionalProperties: false + properties: + code: + description: Python code to execute. + type: string + required: + - code + type: object + strict: true + type: function + - function: + description: Result from RLM agent execution. + name: final_result + parameters: + additionalProperties: false + properties: + answer: + description: The answer to the user's question + type: string + program: + description: The final consolidated program + type: string + required: + - answer + - program + type: object + strict: true + type: function + uri: http://localhost:11434/v1/chat/completions + response: + headers: + content-length: + - '573' + content-type: + - application/json + parsed_body: + choices: + - finish_reason: tool_calls + index: 0 + message: + content: '' + reasoning: We need to call get_chunk via execute_code. + role: assistant + tool_calls: + - function: + arguments: '{"code":"chunk = await get_chunk(''6d291645-590e-498c-a22c-0883e7b1582b'')\nprint(chunk[''content''])\n"}' + name: execute_code + id: call_1o9o5lvl + index: 0 + type: function + created: 1771924598 + id: chatcmpl-401 + model: gpt-oss + object: chat.completion + system_fingerprint: fp_ollama + usage: + completion_tokens: 71 + prompt_tokens: 2356 + total_tokens: 2427 + status: + code: 200 + message: OK +- request: + headers: + accept: + - application/json + accept-encoding: + - gzip, deflate, zstd + connection: + - keep-alive + content-length: + - '11056' + content-type: + - application/json + host: + - localhost:11434 + method: POST + parsed_body: + messages: + - content: |- + You are a Recursive Language Model (RLM) agent that solves complex research questions by writing and executing Python code. + + IMPORTANT: You MUST use the `execute_code` tool to run Python code. The functions described below are ONLY available inside the execute_code tool - you cannot access them any other way. Always execute code to answer questions; do not just describe what code would do. + + CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just call them with `await`: + - results = await search("query") ✓ CORRECT + - from haiku.rag import search ✗ WRONG - will fail + - results = search("query") ✗ WRONG - must use await + + You have access to a sandboxed Python interpreter with these haiku.rag functions (use them directly with `await`, no imports needed): + + ## Available Functions + + ### await search(query, limit=10) -> list[dict] + Search the knowledge base using hybrid search (vector + full-text). + Returns list of dicts with keys: chunk_id, content, document_id, document_title, document_uri, score, page_numbers, headings + + ### await list_documents(limit=10, offset=0) -> list[dict] + List available documents in the knowledge base. + Returns list of dicts with keys: id, title, uri, created_at + + ### await get_document(id_or_title) -> str | None + Get the full text content of a document by ID, title, or URI. + Returns the document content as a string, or None if not found. + + ### await get_chunk(chunk_id) -> dict | None + Get a specific chunk by its ID (from search results). + Returns dict with keys: chunk_id, content, document_id, document_title, headings, page_numbers, labels + Use this to retrieve full chunk details and metadata for citation. + + ### await llm(prompt) -> str + Call an LLM directly with the given prompt. Returns the response as a string. + Use this for classification, summarization, extraction, or any task where you + already have the content and just need LLM reasoning. + + ## Pre-loaded Documents Variable + + If documents were pre-loaded for this session, a `documents` variable is available: + ```python + # documents is a list of dicts with keys: id, title, uri, content + for doc in documents: + print(doc['title'], len(doc['content'])) + ``` + Check if it exists with: `if 'documents' in dir(): ...` + + ## Available Python Features + + The interpreter supports: variables, arithmetic, strings, f-strings, lists, dicts, tuples, sets, loops, conditionals, comprehensions, functions, async/await, `map()`, `sorted()`/`.sort(key=...)`, try/except, and the `json` module. + + Not supported: imports (other than `json`), class definitions, generators/yield, match statements, decorators, `with` statements. + + For pattern matching or text extraction, use string methods (`str.split`, `str.find`, `str.startswith`, `in` operator) or the `llm()` function. + + ## Strategy Guide + + 1. **Explore First**: Start by listing documents or searching to understand what's available. Document names may differ from filenames (e.g., "tbmed593.pdf" might be stored as "TB MED 593" or similar). + 2. **If get_document returns None**: Use `await list_documents()` to see actual document titles, or `await search()` to find relevant content. + 3. **Iterative Refinement**: Run code, examine results, adjust your approach based on what you find. + 4. **Use print() Liberally**: The sandbox captures stdout - print intermediate results to see what you're working with. + 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and data structures. + 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use `await llm()`. + 7. **Cite Your Sources**: Use get_chunk() to retrieve chunk metadata for citations. Track which documents/chunks informed your answer. + + ## Example Patterns + + ### Counting documents matching a condition + ```python + docs = await list_documents(limit=100) + count = 0 + for doc in docs: + content = await get_document(doc['id']) + if content and 'keyword' in content.lower(): + count += 1 + print(f"Found in: {doc['title']}") + print(f"Total: {count}") + ``` + + ### Extracting data with llm() + ```python + numbers = [] + results = await search("financial data", limit=20) + for r in results: + extracted = await llm(f"Extract all dollar amounts from this text as a comma-separated list of numbers (no $ signs): {r['content']}") + for part in extracted.split(','): + part = part.strip().replace(',', '') + if part.isdigit(): + numbers.append(int(part)) + if numbers: + print(f"Average: {sum(numbers) / len(numbers)}") + ``` + + ### Using search results with get_chunk for citations + ```python + results = await search("safety requirements", limit=5) + for r in results: + chunk = await get_chunk(r['chunk_id']) + print(f"From '{chunk['document_title']}', page {chunk['page_numbers']}: {chunk['content'][:100]}") + ``` + + ### Using llm() for classification + ```python + content = await get_document("Q1 Report") + sentiment = await llm(f"Classify the sentiment as positive, negative, or mixed: {content}") + print(sentiment) + ``` + + ## Workflow + + 1. **ALWAYS start by using execute_code** to explore the knowledge base + 2. Run multiple code blocks as needed to gather information + 3. After collecting data, provide your final answer + + ## Output Format + + CRITICAL: Your final response MUST be valid JSON matching this exact schema: + ```json + {"answer": "Your complete answer here as a string", "program": "Your final consolidated program here as a string"} + ``` + + - `answer`: A clear answer to the user's question with key findings and references to specific documents/chunks. + - `program`: A single, self-contained Python program that produces the answer. Consolidate your exploratory code executions into one clean script. + + Do NOT return arbitrary JSON structures. Always use the exact format: {"answer": "...", "program": "..."} + + CRITICAL: You MUST call execute_code at least once before providing your answer. Never give up without trying to execute code first. + role: system + - content: Using the pre-loaded documents variable, tell me when was the company founded and what is their mission? + role: user + - content: null + reasoning: Need check documents variable? + role: assistant + tool_calls: + - function: + arguments: '{"code":"if ''documents'' in dir():\n for doc in documents:\n print(doc[''title''], len(doc[''content'']))\nelse:\n print(''no + documents'')\n"}' + name: execute_code + id: call_2ao52bvz + type: function + - content: '{"code":"if ''documents'' in dir():\n for doc in documents:\n print(doc[''title''], len(doc[''content'']))\nelse:\n print(''no + documents'')\n","stdout":"","stderr":"NameError: name ''dir'' is not defined","success":false}' + role: tool + tool_call_id: call_2ao52bvz + - content: null + reasoning: The environment may not provide dir(); use globals(). Check. + role: assistant + tool_calls: + - function: + arguments: '{"code":"print(globals().keys())\n"}' + name: execute_code + id: call_rjj1hq9z + type: function + - content: '{"code":"print(globals().keys())\n","stdout":"","stderr":"NameError: name ''globals'' is not defined","success":false}' + role: tool + tool_call_id: call_rjj1hq9z + - content: null + reasoning: Probably no python environment? Might need simple print? Might fail. Probably cannot see documents. Could + search. + role: assistant + tool_calls: + - function: + arguments: '{"query":"founded mission company","limit":5}' + name: search + id: call_ee8bn4w7 + type: function + - content: |- + Unknown tool name: 'search'. Available tools: 'final_result', 'execute_code' + + Fix the errors and try again. + role: tool + tool_call_id: call_ee8bn4w7 + - content: null + reasoning: The tool is execute_code. Use that. + role: assistant + tool_calls: + - function: + arguments: '{"code":"results = await search(''founded mission company'', limit=5)\nprint(results)\n"}' + name: execute_code + id: call_o1c87rz5 + type: function + - content: '{"code":"results = await search(''founded mission company'', limit=5)\nprint(results)\n","stdout":"[{''chunk_id'': + ''d43b4445-8425-4434-bc36-e28e6f7e573e'', ''content'': ''Our mission is to make technology accessible to everyone.'', + ''document_id'': ''bb1e588e-8014-4607-bf9b-a077c16d1cdf'', ''document_title'': ''Mission Statement'', ''document_uri'': + None, ''score'': 0.032522473484277725, ''page_numbers'': [], ''headings'': None}, {''chunk_id'': ''6d291645-590e-498c-a22c-0883e7b1582b'', + ''content'': ''The company was founded in 1985 by Jane Smith.'', ''document_id'': ''a2e1e13c-e316-4198-a5a8-49353eaa7337'', + ''document_title'': ''Company History'', ''document_uri'': None, ''score'': 0.032522473484277725, ''page_numbers'': + [], ''headings'': None}]\n","stderr":"","success":true}' + role: tool + tool_call_id: call_o1c87rz5 + - content: null + reasoning: Got info. Need citation get_chunk. + role: assistant + tool_calls: + - function: + arguments: '{"chunk_id":"6d291645-590e-498c-a22c-0883e7b1582b"}' + name: get_chunk + id: call_b0i0vdgg + type: function + - content: |- + Unknown tool name: 'get_chunk'. Available tools: 'final_result', 'execute_code' + + Fix the errors and try again. + role: tool + tool_call_id: call_b0i0vdgg + - content: null + reasoning: We need to call get_chunk via execute_code. + role: assistant + tool_calls: + - function: + arguments: '{"code":"chunk = await get_chunk(''6d291645-590e-498c-a22c-0883e7b1582b'')\nprint(chunk[''content''])\n"}' + name: execute_code + id: call_1o9o5lvl + type: function + - content: '{"code":"chunk = await get_chunk(''6d291645-590e-498c-a22c-0883e7b1582b'')\nprint(chunk[''content''])\n","stdout":"The + company was founded in 1985 by Jane Smith.\n","stderr":"","success":true}' + role: tool + tool_call_id: call_1o9o5lvl + model: gpt-oss + reasoning_effort: low + stream: false + tool_choice: auto + tools: + - function: + description: |- + Execute Python code in a sandboxed interpreter. + + The code has access to haiku.rag functions (search, list_documents, + get_document, get_chunk, llm). + + Use print() to output results. + + Structured result with success status, stdout, and stderr. + + name: execute_code + parameters: + additionalProperties: false + properties: + code: + description: Python code to execute. + type: string + required: + - code + type: object + strict: true + type: function + - function: + description: Result from RLM agent execution. + name: final_result + parameters: + additionalProperties: false + properties: + answer: + description: The answer to the user's question + type: string + program: + description: The final consolidated program + type: string + required: + - answer + - program + type: object + strict: true + type: function + uri: http://localhost:11434/v1/chat/completions + response: + headers: + content-length: + - '550' + content-type: + - application/json + parsed_body: + choices: + - finish_reason: tool_calls + index: 0 + message: + content: '' + reasoning: Similarly mission. + role: assistant + tool_calls: + - function: + arguments: '{"code":"chunk2 = await get_chunk(''d43b4445-8425-4434-bc36-e28e6f7e573e'')\nprint(chunk2[''content''])\n"}' + name: execute_code + id: call_dn2a823n + index: 0 + type: function + created: 1771924600 + id: chatcmpl-408 + model: gpt-oss + object: chat.completion + system_fingerprint: fp_ollama + usage: + completion_tokens: 67 + prompt_tokens: 2502 + total_tokens: 2569 + status: + code: 200 + message: OK +- request: + headers: + accept: + - application/json + accept-encoding: + - gzip, deflate, zstd + connection: + - keep-alive + content-length: + - '11618' + content-type: + - application/json + host: + - localhost:11434 + method: POST + parsed_body: + messages: + - content: |- + You are a Recursive Language Model (RLM) agent that solves complex research questions by writing and executing Python code. + + IMPORTANT: You MUST use the `execute_code` tool to run Python code. The functions described below are ONLY available inside the execute_code tool - you cannot access them any other way. Always execute code to answer questions; do not just describe what code would do. + + CRITICAL: Inside execute_code, these functions are ALREADY available in the namespace. Do NOT import them - just call them with `await`: + - results = await search("query") ✓ CORRECT + - from haiku.rag import search ✗ WRONG - will fail + - results = search("query") ✗ WRONG - must use await + + You have access to a sandboxed Python interpreter with these haiku.rag functions (use them directly with `await`, no imports needed): + + ## Available Functions + + ### await search(query, limit=10) -> list[dict] + Search the knowledge base using hybrid search (vector + full-text). + Returns list of dicts with keys: chunk_id, content, document_id, document_title, document_uri, score, page_numbers, headings + + ### await list_documents(limit=10, offset=0) -> list[dict] + List available documents in the knowledge base. + Returns list of dicts with keys: id, title, uri, created_at + + ### await get_document(id_or_title) -> str | None + Get the full text content of a document by ID, title, or URI. + Returns the document content as a string, or None if not found. + + ### await get_chunk(chunk_id) -> dict | None + Get a specific chunk by its ID (from search results). + Returns dict with keys: chunk_id, content, document_id, document_title, headings, page_numbers, labels + Use this to retrieve full chunk details and metadata for citation. + + ### await llm(prompt) -> str + Call an LLM directly with the given prompt. Returns the response as a string. + Use this for classification, summarization, extraction, or any task where you + already have the content and just need LLM reasoning. + + ## Pre-loaded Documents Variable + + If documents were pre-loaded for this session, a `documents` variable is available: + ```python + # documents is a list of dicts with keys: id, title, uri, content + for doc in documents: + print(doc['title'], len(doc['content'])) + ``` + Check if it exists with: `if 'documents' in dir(): ...` + + ## Available Python Features + + The interpreter supports: variables, arithmetic, strings, f-strings, lists, dicts, tuples, sets, loops, conditionals, comprehensions, functions, async/await, `map()`, `sorted()`/`.sort(key=...)`, try/except, and the `json` module. + + Not supported: imports (other than `json`), class definitions, generators/yield, match statements, decorators, `with` statements. + + For pattern matching or text extraction, use string methods (`str.split`, `str.find`, `str.startswith`, `in` operator) or the `llm()` function. + + ## Strategy Guide + + 1. **Explore First**: Start by listing documents or searching to understand what's available. Document names may differ from filenames (e.g., "tbmed593.pdf" might be stored as "TB MED 593" or similar). + 2. **If get_document returns None**: Use `await list_documents()` to see actual document titles, or `await search()` to find relevant content. + 3. **Iterative Refinement**: Run code, examine results, adjust your approach based on what you find. + 4. **Use print() Liberally**: The sandbox captures stdout - print intermediate results to see what you're working with. + 5. **Aggregate with Code**: For counting, averaging, or comparing across documents, write loops and data structures. + 6. **Use llm() for Classification/Extraction**: When you need to classify, summarize, or extract structured data from content you already have, use `await llm()`. + 7. **Cite Your Sources**: Use get_chunk() to retrieve chunk metadata for citations. Track which documents/chunks informed your answer. + + ## Example Patterns + + ### Counting documents matching a condition + ```python + docs = await list_documents(limit=100) + count = 0 + for doc in docs: + content = await get_document(doc['id']) + if content and 'keyword' in content.lower(): + count += 1 + print(f"Found in: {doc['title']}") + print(f"Total: {count}") + ``` + + ### Extracting data with llm() + ```python + numbers = [] + results = await search("financial data", limit=20) + for r in results: + extracted = await llm(f"Extract all dollar amounts from this text as a comma-separated list of numbers (no $ signs): {r['content']}") + for part in extracted.split(','): + part = part.strip().replace(',', '') + if part.isdigit(): + numbers.append(int(part)) + if numbers: + print(f"Average: {sum(numbers) / len(numbers)}") + ``` + + ### Using search results with get_chunk for citations + ```python + results = await search("safety requirements", limit=5) + for r in results: + chunk = await get_chunk(r['chunk_id']) + print(f"From '{chunk['document_title']}', page {chunk['page_numbers']}: {chunk['content'][:100]}") + ``` + + ### Using llm() for classification + ```python + content = await get_document("Q1 Report") + sentiment = await llm(f"Classify the sentiment as positive, negative, or mixed: {content}") + print(sentiment) + ``` + + ## Workflow + + 1. **ALWAYS start by using execute_code** to explore the knowledge base + 2. Run multiple code blocks as needed to gather information + 3. After collecting data, provide your final answer + + ## Output Format + + CRITICAL: Your final response MUST be valid JSON matching this exact schema: + ```json + {"answer": "Your complete answer here as a string", "program": "Your final consolidated program here as a string"} + ``` + + - `answer`: A clear answer to the user's question with key findings and references to specific documents/chunks. + - `program`: A single, self-contained Python program that produces the answer. Consolidate your exploratory code executions into one clean script. + + Do NOT return arbitrary JSON structures. Always use the exact format: {"answer": "...", "program": "..."} + + CRITICAL: You MUST call execute_code at least once before providing your answer. Never give up without trying to execute code first. + role: system + - content: Using the pre-loaded documents variable, tell me when was the company founded and what is their mission? + role: user + - content: null + reasoning: Need check documents variable? + role: assistant + tool_calls: + - function: + arguments: '{"code":"if ''documents'' in dir():\n for doc in documents:\n print(doc[''title''], len(doc[''content'']))\nelse:\n print(''no + documents'')\n"}' + name: execute_code + id: call_2ao52bvz + type: function + - content: '{"code":"if ''documents'' in dir():\n for doc in documents:\n print(doc[''title''], len(doc[''content'']))\nelse:\n print(''no + documents'')\n","stdout":"","stderr":"NameError: name ''dir'' is not defined","success":false}' + role: tool + tool_call_id: call_2ao52bvz + - content: null + reasoning: The environment may not provide dir(); use globals(). Check. + role: assistant + tool_calls: + - function: + arguments: '{"code":"print(globals().keys())\n"}' + name: execute_code + id: call_rjj1hq9z + type: function + - content: '{"code":"print(globals().keys())\n","stdout":"","stderr":"NameError: name ''globals'' is not defined","success":false}' + role: tool + tool_call_id: call_rjj1hq9z + - content: null + reasoning: Probably no python environment? Might need simple print? Might fail. Probably cannot see documents. Could + search. + role: assistant + tool_calls: + - function: + arguments: '{"query":"founded mission company","limit":5}' + name: search + id: call_ee8bn4w7 + type: function + - content: |- + Unknown tool name: 'search'. Available tools: 'final_result', 'execute_code' + + Fix the errors and try again. + role: tool + tool_call_id: call_ee8bn4w7 + - content: null + reasoning: The tool is execute_code. Use that. + role: assistant + tool_calls: + - function: + arguments: '{"code":"results = await search(''founded mission company'', limit=5)\nprint(results)\n"}' + name: execute_code + id: call_o1c87rz5 + type: function + - content: '{"code":"results = await search(''founded mission company'', limit=5)\nprint(results)\n","stdout":"[{''chunk_id'': + ''d43b4445-8425-4434-bc36-e28e6f7e573e'', ''content'': ''Our mission is to make technology accessible to everyone.'', + ''document_id'': ''bb1e588e-8014-4607-bf9b-a077c16d1cdf'', ''document_title'': ''Mission Statement'', ''document_uri'': + None, ''score'': 0.032522473484277725, ''page_numbers'': [], ''headings'': None}, {''chunk_id'': ''6d291645-590e-498c-a22c-0883e7b1582b'', + ''content'': ''The company was founded in 1985 by Jane Smith.'', ''document_id'': ''a2e1e13c-e316-4198-a5a8-49353eaa7337'', + ''document_title'': ''Company History'', ''document_uri'': None, ''score'': 0.032522473484277725, ''page_numbers'': + [], ''headings'': None}]\n","stderr":"","success":true}' + role: tool + tool_call_id: call_o1c87rz5 + - content: null + reasoning: Got info. Need citation get_chunk. + role: assistant + tool_calls: + - function: + arguments: '{"chunk_id":"6d291645-590e-498c-a22c-0883e7b1582b"}' + name: get_chunk + id: call_b0i0vdgg + type: function + - content: |- + Unknown tool name: 'get_chunk'. Available tools: 'final_result', 'execute_code' + + Fix the errors and try again. + role: tool + tool_call_id: call_b0i0vdgg + - content: null + reasoning: We need to call get_chunk via execute_code. + role: assistant + tool_calls: + - function: + arguments: '{"code":"chunk = await get_chunk(''6d291645-590e-498c-a22c-0883e7b1582b'')\nprint(chunk[''content''])\n"}' + name: execute_code + id: call_1o9o5lvl + type: function + - content: '{"code":"chunk = await get_chunk(''6d291645-590e-498c-a22c-0883e7b1582b'')\nprint(chunk[''content''])\n","stdout":"The + company was founded in 1985 by Jane Smith.\n","stderr":"","success":true}' + role: tool + tool_call_id: call_1o9o5lvl + - content: null + reasoning: Similarly mission. + role: assistant + tool_calls: + - function: + arguments: '{"code":"chunk2 = await get_chunk(''d43b4445-8425-4434-bc36-e28e6f7e573e'')\nprint(chunk2[''content''])\n"}' + name: execute_code + id: call_dn2a823n + type: function + - content: '{"code":"chunk2 = await get_chunk(''d43b4445-8425-4434-bc36-e28e6f7e573e'')\nprint(chunk2[''content''])\n","stdout":"Our + mission is to make technology accessible to everyone.\n","stderr":"","success":true}' + role: tool + tool_call_id: call_dn2a823n + model: gpt-oss + reasoning_effort: low + stream: false + tool_choice: auto + tools: + - function: + description: |- + Execute Python code in a sandboxed interpreter. + + The code has access to haiku.rag functions (search, list_documents, + get_document, get_chunk, llm). + + Use print() to output results. + + Structured result with success status, stdout, and stderr. + + name: execute_code + parameters: + additionalProperties: false + properties: + code: + description: Python code to execute. + type: string + required: + - code + type: object + strict: true + type: function + - function: + description: Result from RLM agent execution. + name: final_result + parameters: + additionalProperties: false + properties: + answer: + description: The answer to the user's question + type: string + program: + description: The final consolidated program + type: string + required: + - answer + - program + type: object + strict: true + type: function + uri: http://localhost:11434/v1/chat/completions + response: + headers: + content-length: + - '1388' content-type: - application/json parsed_body: @@ -874,24 +2313,26 @@ interactions: - finish_reason: stop index: 0 message: - content: '{"answer":"The company was founded in 1985. Its mission is to make technology accessible to everyone.","program":"# - Get relevant documents via search\nresults = search(''company founded mission'', limit=10)\n# Find founded year - and mission statement\nfounded_year = None\nmission = None\nfor r in results:\n text = r[''content'']\n if - ''founded'' in text.lower() and founded_year is None:\n import re\n m = re.search(r''\\b(19\\d{2}|20\\d{2})\\b'', - text)\n if m:\n founded_year = m.group(1)\n if ''mission'' in text.lower() and mission is - None:\n # simple extraction of the sentence containing \"mission\"\n sentences = re.split(r''(?<=[.!?])\\s+'', - text)\n for s in sentences:\n if ''mission'' in s.lower():\n mission = s.strip()\n break\n# - Output\nprint(f\"Founded year: {founded_year}\")\nprint(f\"Mission: {mission}\")"}' + content: '{"answer":"The company was founded in 1985 by Jane Smith. Its mission is to make technology accessible + to everyone.","program":"# Consolidated program to extract founding year and mission\nimport json\n\nasync def + main():\n # Search for founding year and mission statements\n results = await search(''founded mission company'', + limit=10)\n founded = None\n mission = None\n for r in results:\n if ''founded'' in r[''content''].lower() + and not founded:\n founded = r[''content'']\n if ''mission'' in r[''content''].lower() and not + mission:\n mission = r[''content'']\n # Fallback to chunk extraction if needed\n if not founded:\n chunk + = await get_chunk(''6d291645-590e-498c-a22c-0883e7b1582b'')\n founded = chunk[''content'']\n if not + mission:\n chunk2 = await get_chunk(''d43b4445-8425-4434-bc36-e28e6f7e573e'')\n mission = chunk2[''content'']\n # + Print results\n print(''Founded:'', founded)\n print(''Mission:'', mission)\n\n# Run the async main function\nimport + asyncio\nasyncio.run(main())"}' role: assistant - created: 1770373471 - id: chatcmpl-972 + created: 1771924607 + id: chatcmpl-899 model: gpt-oss object: chat.completion system_fingerprint: fp_ollama usage: - completion_tokens: 246 - prompt_tokens: 2135 - total_tokens: 2381 + completion_tokens: 296 + prompt_tokens: 2645 + total_tokens: 2941 status: code: 200 message: OK diff --git a/tests/cassettes/test_sandbox/TestDockerSandboxContextFilter.test_filter_applied_to_list_documents.yaml b/tests/cassettes/test_sandbox/TestSandboxContextFilter.test_filter_applied_to_list_documents.yaml similarity index 100% rename from tests/cassettes/test_sandbox/TestDockerSandboxContextFilter.test_filter_applied_to_list_documents.yaml rename to tests/cassettes/test_sandbox/TestSandboxContextFilter.test_filter_applied_to_list_documents.yaml diff --git a/tests/cassettes/test_sandbox/TestSandboxDoclingDocument.test_returns_dict_for_document_with_docling_data.yaml b/tests/cassettes/test_sandbox/TestSandboxDoclingDocument.test_returns_dict_for_document_with_docling_data.yaml new file mode 100644 index 00000000..29228880 --- /dev/null +++ b/tests/cassettes/test_sandbox/TestSandboxDoclingDocument.test_returns_dict_for_document_with_docling_data.yaml @@ -0,0 +1,42 @@ +interactions: +- request: + headers: + accept: + - application/json + accept-encoding: + - gzip, deflate, zstd + connection: + - keep-alive + content-length: + - '95' + content-type: + - application/json + host: + - localhost:11434 + method: POST + parsed_body: + encoding_format: base64 + input: + - Docling processed content + model: qwen3-embedding:4b + uri: http://localhost:11434/v1/embeddings + response: + headers: + content-type: + - application/json + transfer-encoding: + - chunked + parsed_body: + data: + - embedding: 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 + index: 0 + object: embedding + model: qwen3-embedding:4b + object: list + usage: + prompt_tokens: 5 + total_tokens: 5 + status: + code: 200 + message: OK +version: 1 diff --git a/tests/cassettes/test_sandbox/TestSandboxHaikuRAG.test_get_chunk.yaml b/tests/cassettes/test_sandbox/TestSandboxHaikuRAG.test_get_chunk.yaml new file mode 100644 index 00000000..6977ce26 --- /dev/null +++ b/tests/cassettes/test_sandbox/TestSandboxHaikuRAG.test_get_chunk.yaml @@ -0,0 +1,82 @@ +interactions: +- request: + headers: + accept: + - application/json + accept-encoding: + - gzip, deflate, zstd + connection: + - keep-alive + content-length: + - '99' + content-type: + - application/json + host: + - localhost:11434 + method: POST + parsed_body: + encoding_format: base64 + input: + - Content about foxes and dogs. + model: qwen3-embedding:4b + uri: http://localhost:11434/v1/embeddings + response: + headers: + content-type: + - application/json + transfer-encoding: + - chunked + parsed_body: + data: + - embedding: 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 + index: 0 + object: embedding + model: qwen3-embedding:4b + object: list + usage: + prompt_tokens: 8 + total_tokens: 8 + status: + code: 200 + message: OK +- request: + headers: + accept: + - application/json + accept-encoding: + - gzip, deflate, zstd + connection: + - keep-alive + content-length: + - '75' + content-type: + - application/json + host: + - localhost:11434 + method: POST + parsed_body: + encoding_format: base64 + input: + - foxes + model: qwen3-embedding:4b + uri: http://localhost:11434/v1/embeddings + response: + headers: + content-type: + - application/json + transfer-encoding: + - chunked + parsed_body: + data: + - embedding: Gk9OOB4t5jwFfxg8FOqBvNCJhTielhw92bZ0PSkI9LxgixE9OLMKO9adDL2Ol448phCiOzV5K70Q/O484o0/PDsHHT185Eu9ZuAwvZEP97vBOOy8rLy/OkD4jLyHdJU7e/6GPDoIvrtQf8a8lBmxvI5ShjwYIRC8swkpvYHmsrzt4TU9hw70Ou6xxjvSggy9ttAHvKktDrxNbzy89wIuveEBtjvfJSi9aK+sPDkpqbpmdZ08HVptvDFhhzv2kUu89zy8vDAmx7zpiZY7bXAUO5Gz3DwDcJW8SLcSPFfx6ryHG5Y9jzcavNjZOzunSfA8FAULvLSKwbyeygi9KL+BvBNOy7ujlJ+8FHCgvJ2PwbyDXqs82vGOu6egW7z7ki89R0ndvFDUnLudqxw87hsFveP+CbxiB5s87MHCPJo7mTzEbZQ85nVHPEdugjxgmwQ9CQM+PdjxJDyg5xs9NitFOgE/F7zkET88KfdhPEzUBT3zqSw6ez9mvDw1A7y1YaE8hsRtvHK+dLyXLL68xsSpPL7qc7w+yyO8GDEaPcIm0LsH07S8RfDmvA5BtLzrgX271JjPOJkMkbxWJt27M2CUu+TH8DzKUsu8UVpCvPtJlLvA2M071qxjPTG6HTyTv8S7oBhqvFWGxDwDg1U7NXn2PFb+tLnVnkK87r4IvNyplbwiSca8q0CJPGGIAD0Md+u75g8VPCVRtjq42LA8BUiZO4IIgDvtGaK8xuQvvMgVk7nFjdq7JuUou3dgjjuTGoQ8nJfHvJWgQDwYQ028DtJxuyjSjjwUVjg6xn/fPBEUO7oUn/w7O7G2PM1B2rk2RWk8d4OevHUa87nSZNA7PwtFPBHd9TsOyNM856+4O7vP5zyGKMM8cfMzuqmOcb3OxtS7gBtevHC6tbw8fNy7NcvpvHHhTTyI3Ia8+HiFvCiYfLy839e8CSMBPE2pCzxiaeU7zEl0O3NRSDyWc527267TOerGkTxMqiy7ZYJiOi39gTzojwI8p2ywO7q2Ur1dMl88XFHJOh+2oLwNC8w7FPp+u7HTB7wB5nA8JhievIBBIT3Dwgq8wP9nvHATaLwlIHy7udlJPGiOiTuGxwy7kjiivMcAVjuKXHA7RldaPDCEl7wBnLC85aTLvCnJVbt0jUW8McG7vNm1krxpJTw9A3rRPIJ9Ebtg+DM8LBQ5vNPTTzn9BAq9p3+4OxqFMjzi2yY8axatOo2JCrzDzUK7SpyGPPrPpTr/Gv27Tv5avFI2hDyOtNY8APkIvCSZgbnmfjA89fcDvVKZq7zGhSi8pV/CPGIjuTvMzfy8XskBPJ9Zdrumi3K8iYevvEgmabxkKR28iRWNPJ/Jirx9Gtc5kwXQu4O0r7wDEbG89PuavO30rDw9CG06gbNtPCOuKLyCzU07INMzPJiHD7xHf2k8gFVovAVEhDwGKYu8XceUPTtcdzvNBsI88J9KPFrHSboDxDs7QkmiuO18O7pzqsI7w3MTPVo99rviI1e8iCpFubBbdjz+BJc8fIEgO2+6HTnobow8/3g/vLK5jjmGIkk8GQmzvAG99TzrNIy8dPxqOrwb3TxdbZc7E5w2vLWQv7s3WdK8Bv5ivJkQTrzNYag7aVVAO8lbFDtmRME8DBQJvKOqW7tXgQa9xdi4vBN51Lth0MS8G4HHOnn3TTwsV6Q8X3Z1vQgqYTs/wfQ7yG2CvDhhEb2+akU802snvcx1QLx7Rf+68CpcvC/8RzoAeOE7dosEPAmRB7xnR3A9oXkBvBEQ7TzmkUS8sNyrvBgsALwtAwC83iFpuiHJyzyoayw9UV4zPAq5lbxBDo27eHrSPLhVHLwuqMW711o1PI8isjxMqQ88zRCqvL7HGb1NeU08QZ7HvOQeJrt4jTK8DLVVutw43DuKvmC7nqmQO3VOJrvexfS82dh8vA/P3jhYRZ07O/2cO3+A9TsYlB47x4KyO9TPDT3tBV29BsesvNtO27tbVyO8ZaXWPFAeZrydHws9OERovObR3jx/XiC9XzJsvGsoxzuoWb48RC9TPBHZhDupTLo7gzUcuypzULvG+ky8UVEcPFOo2LzhpTq8a4XcunCyWDz7c4M5D9IMPYmBnDtL4pQ8zWNIPAIGhTwqzLI8WxAYuuu+VLz1Bne8JZE0vRw8Cb337t87er4TPLfbXbyfpo89rkkIPA3Szbvd2t68xVP+PHgKoTydZJU7MJgQPFMntDzEoI886bq0Oz1odbzzmpw8OQ7RO/YKDbuKti67NwaxuZ++Aj17o2Y8U6GpO78vZDzbi4m8HJMkvKy6HLyebXK86xaYPPwGPT0YpZK7zML6u3O5GrtT8/e8uBofPOlMgrsdOMi7xgpdu28QazzaoeU77ur8O+J65ru2RLc8N3kvvAHMxzy71Uu9699PvBIYkrsgAg88kbs7Ow9g/7zr5kq7aV6nOejphrwB66i7UUUluqkxwr11df47kAXouwsmprpi1Zs6c+jKvEr+drx1O668GeKpPGTwGT3K7ym9LTQIveJftLwms6E8cO/LvAYPIbwUvGM8+OuDPFAuXTzZl147POl2PMjIjju4eGe8XvjSPJuBIruup988KnNKPBrFHj23wYC8oq3OvElfqzy6fKw8ee6QO6AAsDvql4c8Dw6LPG8zhzzlBhq9SJ8MO/zNezwnWTi8iofyPIsnezui8re8SVG9vPv5pzx3oRM92Ok8vIZPkDwS6P48lFqUOw7RdzwIbgc8m9LWvIkapzz8P0e8ipegvJGPnrvxyZY83NJwvG7x6bxQIu27fQPJPGlK9ryFwuO88Jm0PKo+2zxxwrW8luJDPHKQWri35VC8idIWvOpzf7wMJpO8VIXjvNgRHzzQqwY9GDo1PW3YjDwXlba7OYQEPd76irzTxsY85ZMKPFWaRTwHIIU7JIIFPMzttDwWxNe8Rx7mPFTXNjuqHjm8kbdiO3MUv7x0F3s78rd6PFBJMjxca987V0U4vJu1bTwy/SS7RXGePLfCrbwemmQ8EKUQPIOtPL2/5s+8eCv+vFenybyZd168whYCPBIYjTz3FuU8Ez4HPK5tFL2FbDW8D/iyPBNYq7wYeTy98RkcPFvtHjyp5ow8Z9m4PBXK7bzPB5s8XMZbPN8TWj3kCGg8BgXIPKCyDLx34F+798qlOviGS7smFkm9nGCTPAsyQLyzVhw8ze6PvC1KwjuMT6G89UyHPUoRvLz6DIm8J/DQOnxdKbzkP7O7C9ZEPOnxzTxWVYe8kxr3PE5isLzfZ9m8r+qovIM6ijzLa+Q840BUvcFZe7xl1Pq8ORCNvGfXrbzQpMy8M4vVvHEixzulppY8Gr3HPB1MrL3bwvQ7BKNjPHag/TlFzXi8p406vNooorxaaTs9M1dUvC+o7jzzVRW8rzxPvPHeHD2Pu4g88zPHvJwUrTyeCGK8jQItu/uENTxXoBE9iRwIvGcLp7zqiIW8TIdDvKKfwTxu+E88ZUQTu8CxALy4B0Q8Dpj6PIffSzza08U86x3dO97AyTzyNta7jdq9Oy/uNryUntY8pNW3PI55mLsBfYc8VLK0vOQqE7xwgw49beUDO4zAqzyxMwM9giIYvW4PhbwAfbU7Sf8jvY/0NLzu2aO8NKnBOa+/EjwMgs+81W+AvKkmHD1R15O8vv3auwxoWLw6RDk8+OLWOzRXrTyAVrg8xAzIPIZIVLxB0ri8OF8CPSRtBDo07pw6V0xNPC3I3zz2T6O8VBmrvGPGhLu2v/48htDbPExxGzttmnw6e3cCPXp3zTyyO5I8Kd6Su+KAFD1n4JO7t4GlvGC3nrv6MVa898AVPAAllrws3iq9a6WavE9YJzyCxWe8pTUdvbulqTzB/QK9LQp4O5CAXbzGy4A8p4sXPO2Cojsl3cy8xPnEPDHvCD3iQ5e6y/eDvEaD5zwRQJ68WJRTuwXbTjxB1QI8pgMpPE5lTrygF1s8p3MmO9OcRjxf/7A8lFQtveSnHz2qz0W9nP21vDh5NDzWezc84Ti+PG6eSzzzpw27LnmdvHFsOTpeytS6SzwNPFyutTybnQY9d2qXPMdW/bt+DXm8FO2JOyS9eLwPLZW8tsF+usFsUjwO1V+8KpaIvBeSnTzxtcm8rmbvPGNIx7ytNaa8Qj/KPHIehbq5fri8fN6LvBMAMz36kIe732zQPFsPXrwGwKs6IJxTOYxh/7w20xC8yVePvGcug7zd7L67nhjHvEYrnjzozyw9RjFOOSGQkrz7nac8TJLpvIIoRrv+wRm8Q0LBvDKfE7zldwy84oJSvC49DLxU5XA96HdlOy3xkjwRwP07t3eVPOrDm7vqNNs8UoioO/JKzbyfiXo7y2efPBmStrsb/gi8fs/cPObbcLyG2MI7ZwLfPJ/04jt/Bgo937amvPBThTqgh408pWbRuz6Vdziy9q28XgGwuqSszrtWhRc7+LrWu6jZNDzERA088SaTvG7wRbtbyjc6+v8WvdpsIL1iWAw6qJimPEYBNjzVmbc8zBuavLWZibvciO28SN77uykwArxQbkS7VpTOPFWHITogkFM9w3g5vLo8hrx7jz07AaeSPLAv+TyDO327wqurOxWBxrw6yIM8NCgpvXa2JrveJ/W8L20tu6y6zztpurK8smcxvP6xhDuvv/C8VxdaPHSPLzz8Bqc8TLfyOs0qlTxf71a80qXKPD2o07uBTA07vSG+u+9pCDyXSx+7ArRLPduXqDxe+Ju8flysPAmElbw64wC8KK2su+HpmruC5te6KAeAO9QvlrzLIlG8PfdkvMN+2LvIJ8Y8OoPDvOTTUzy7BuG7iYKsvM3oxTyGmC08hOezPCaQELwjZyc8tj2kvE9PPrwHVFY8ixsxvBByMrs1vfe8oBlCOss4C7uXrFW9IBv/O9ySl7uaeNc747RRvDcYvbvycgI7aEWZvH3NVr3jHfW7eZy6PeWRBr2ftA+6JnbzOqoIJL0Ytkm713wUvQrYr7saa6S7mlkavCX0pLzmy8282yMrvEtPiLtXOMY8o1CjOwuKGbtqtfE7O+WEu6t/cD2qiP68V7RIOpVwUTzibac6jMa6vOaW7DvAFyO7BtIfPPoUtLzAoXY8j1MEPNPuyLwOdKy7LTKWPPC7Izzq3Fm8JxEBvFpGUzxq6m68AsiMOyZ5Dzx6LlM8dlCsuwkBsLv9JAS8PjQnvMpD7Drjngy8wVmAPKdX6Tt8fZs7sjabPE48h7xCoBs9BXEJPaIVOj1H0De9aptZOw1OrDtKUI28DpTdvPOVkbw1O7O7FdO/OkhrrLqgOx08y0JnvIFN8zsuRNw7UDAfvPccVLuEhOs8ZUqXvLRsvDsYiac8EBs5PMZ0Ir0nWMI86LZIPHddMrwBtYo7hfvXvAMjLjxoq8O8uRH3PHTd4Lt+T8a8GNwNO9++Er0B8Gm8bcmROVMheTtkWBm8CWIuvNXa5by6rpo8m3GvulqnCbo5D/A78eFLvFOZljobuc27pt4cPef2sLyf2hS9LkMCPGxHGrxHZqK7FL+EO1xCuDtx1xS8XglSPF4tervbGZA8jRYVu/Yb6DtrRjy87smVuwmlkLxOUIi8PxW2u8wMFLystmU8caUru0JJWbyqv/m7Ks0ju4ZwSDwh9DS8uIAXO0azHztk06U7u4UaPABoEryXvE45yzgyPJ3XuTzuSrK8bdu4u/8hBDzp6ec65ZMLu1X+pzzoKcG88/wrvH80yzyaOLK8aWziPIyyZLzsLBG6kao/O/pDzruv5ho93paTvEfrw7umEL+89tr/vEWHzLtB3f68ZUmvvH6hR7zukzq8W25MPV7QJ7si7OW63DDaPOUmajwp3pm838AKvKCaPbufVbg8WcqtvMLejLu/B+o76cSbPB+3ObztSCM9SUUfPHG6/bvBY1O911y0u+8Yg7xe/Fk7WQ3lPEw6Yjx+zwc9lzJSvIDhjzzaiU685bahPEuklLuGSpY7XuoDuwhC3Lrt2nu8L3+au5A9SLzpJZE8JKKvu7RQSDwlrSm8z9KJvI9UMTxHPyC8Pmo+vQ7xWbu6YjC8u8FSPJ5kYjzynzm7OctPPRdoQDwZGba81fBmPP2m0Ty98Ak8PicavJzgJTxfHEe8o5+6O4WSHTy58ty7RvLGuhoh0jzN7ag7TyQEu1SIIr3X1As70A8CukkBcjztpYW8bBQwO9OkmDv7lW88PUm8O1kGuTy54Ie6SS2luxPnJjqtOpE8jp4fvXA44Tydmi28JqqMPEQunbsTSBa9oH6fPOyBSDsiaY88aGq+PK2tBbyneT89wl/3PBgNrjukyG68MVlJvTiXkzvo4GK7bwtUPP00vbvousq8ubO9O25Ax7tjiSs8WF53vACfJTzu5zE8JKRuvBMfgbtMIUS8m/KDPO+E6buI7Qg856j2vJDft7ttvua8n9UhvCDVG72EKAE8ZvvfvEPQLbzuS9+7mJQZOxhRqDzZUuE7cTgsPZrmw7y2kWO7b8ulOzQqpDz3YZq8Jpu6uiUTDD3qBcy7aQ0DPVSHgLo+8Cy8ulHZu31Xh7oPpai88z+qvHw/JzyKwYW6uFPwPNMMpzwZ2aq8Kz0DPN9JOTwFmlo7V8AQPWcgO70w0Vm7edsEu85isrzl6Dg9kr6qPIjmITowUqU8KR4YvBbKuTziKX89PPtdvBtPpbsfbms8CC0RPE3sDzxde6m8//EsuwNpeDvkIV67WZg/PFRUo7y4XZW88LY2u+R5tzwRmrk8mMWDvJxsO7xVbpa8Y8p9vC5Xcbw3hBs75JNgPK0h2bm4XDA8jU1PPIUG3juDXNw8FDySvEkZ/rrTW6i8VsAIvNGSvrt+Wya9AHh3uugr4LyZHXo8QPIlPbf2FjzRe7y7NDo7vV+Q1rwfaWi7cugcPXdK97txfSm9MRHQPCqjh7yvWKi8iHiovKqSHrwe9yS907DYPNs8DrzUP5a74w4CvfnX1rtLTUm8cXMSO8fMgTtoR2o8mPMQvQXAoTysukM7o+fGvFx7pbyHb4Q7CkuXOxUlZbvP3hW8QFwPPcKZTLxEy6G8NXqLvAy0Nzwn7IW8O1utuxZhGrw6no88+rqJPC4aVjwfqfa8QQtSPKxcerxMaqu6uqbiOxfeFLymVXY8XsjcvDnI4jx774U76fsevF6aDT2q+Vs86nAOPTamtLxSQMa8Meh9uu06PrwaN448ZoGqPKpmnjsw/i+93v2aPL+c3TxM2s283WKsPBGk/jvH64c8MXa1PEBPBD3CY0S7lbfyPFMQLzsJNN07NaEMPLgbdTwIZt+8ZaskPP8gqLyK+DW8nZvJPNkBjjv7t6w7oWCEPFHr0zwFkZU8YEkKPVq4NLxqaOw7xUo8O/v7PDuYe4m85YavOmESu7z/Pn68gapMPEtxC73GeuS7VMwEPD4iITzgu+87mF4RvacXWDxIJgE9KUyoPERSU7zzRH27NgHYvFwG9zxtM8m7vFY8O3GFxjxtjtG79HonvMVAbTwr3487zwkJPBpmb7x9+nI83k8uPGSElrotzp48z+24PHkOarz34jW9mdAzPBlCTj0Wlj08SZpHvWvC3rw1l8K7wrMNutdDjLsCpFO8+F9Tu5ZXFD2Amwc8eJb8vEIosTvhDwG8sg2PvNaJwTzPhrw8SC5Au04PXbvwCJc8SUiZu61ilbrIHp68aV4kPAWTUzySmiA96Lq2PJI5SD1mowk8gQUJvXtjxLyBu/080EwUvIhRqLzgdXm8rQ3+O6WSfDyhCxW7SrZePOSUJL032/o7LAglPWUc6Lt2Gkk71N52Oo4oAz3aQX67JiSGPHzphrs777a6kM4oPKYVLjyBsAo8Zv0TPVpTobvxCGW8Zx8XO7WpMjyj1DK9vO6Mu5b+aDwqJBs8QqBqvPaZDjsQ+Sc9mff4PILj3LtLGi+6I374vMsoKTwwQA49hUEBukmq6TxVE2q9IxNhPPH4PztU4l282A0DPH6YWDzs3QY8lb+du/HbnLs2GII6AFwkPdQrSryb3QO9v4uAvN2n67yKTRO9hE+QPNcwiLx8Y3C8uFmLPANg57yPIfi8GAAvPQXEoLzTrK48AzIzPLnSXbyVpvu7mFdOO1ZTujwgR0G80TKnPC7LHryyeZg8b2W8vOAubDsaoMI7jLIyvfNqnLsL/Ie83WW5PPslWTpxUDs8tJ94PAoxMTwMkGg88BVsu1fmtjqExrA8uQ3/vCigWrwWvcc7AaYsvDjP9Tw701+8Ht+rOw/vQrzB6X485TGvO7E8ujw/ykk8VgZFvKE3C7ug51g8IFpHu2aELzvdrie8YzqIPMr9KjtTL6i7h6EQvRiZuTxoqD47iGEgO8fNbbxrjcY7iIJuvCWxLT0+oq+8F+ocPDsOPzsWtRG8UXV8vPneyjybotm7cGtOPGP4B72EkpQ8OT/qO3Okpzwk11e81z6zO4IflDz+gPQ7M5ZnvLAeQjy81Y473ByivAJ9ijsJvhU7MmwCPKJDWzyg2oe8ZcyzOwKWPTxsDcu8FY/EO70E2TszlkQ8q5NzvAayrbudLcG8BWsfPDjRnzzSRLi71RRyPKuKcropFMG8nsetu5s+8jwqHQS9DMqBvIq2zjxMklq8SPCCvFLtLLo25tM7btzfO5pDgTzvH628HEE2PHZtZrwQGMW7N5RmPITTVTyjYua7WGIRPX0uyDurEKu6sEvbvFfkED0xADk9HCKjPEBfpDysAwC9hYnbvKxotzx9gN+76xaAPNDiz7sbBMq889TXus6H17zidBY9MB/UvJA0srvRCwy6wLWkvFIkHTwTKiC937uJvPIkOTtgz6076I/uO0eIeDweE8w71SIRPZHvezt+0EY6enUDvD+1NzzdLtw7y/lnuy0mvjoAXjg771UbO71/hLzoymm846BrPXlmzzxIIGS8v2SUOX2FmbyT8EW8HmjxvAwDjDzQP/y6gxPTvOAI1zyDFJc8rNsNvJDR+Lyj1q68apYRPffHKL2lOTy8CgNlvEc1mby+dxG9fZLNPL3muTwdGuO8fCYtPLAeobtmOIa7VqwVOxlFvzsZTjy87vWWvPY1UD2YQiC8Bdp1uoPu6LwphrG830RsvJ7/njyQOtO8vSaIPO28qDl1Zks7p/CRPBJDgjxMbQg8cbZYO0ok97zMaK87EhcuO+kfWTwGAxS8/o/lvHs4+jtByFg88VOrvPaZxrxMi4Q88b8APdJmRTxhmpQ8NzKtvKVQ7bs7nqI6/32cPJ9jET2MpNw8ZSAlPH0ACD3VcaM85F+xvEH4TDvB1im9rZOLPJyvy7zLZnI6jEdrPE0TnLwlPSG8mzimvFqp/zxSQgS9BqcPO9aIJzvSOqW84z7/O4LIoDyIfCo8JlNtvLsfgTtgvTY95eEDPOcKEb1xGdW83mbePEbxvDw9mB08kDwPu9DlbDo0iKC7C+w5vDF2xrzW3le9It6FvNMbDDzruWS7f4iLPHmgDLyDeLg5t3LMu/F2Gj0hDrE79EEkvPN4kDwWrpC8TPidvCnCoTxe2+u8cQEWvOkzzrvrmDs8J4jLvP9aZzzBjLO7sHkaPXiR8zu55Bc81kdAOiQxursPfuG8XYWau/b69Lkt3wE71EeLuneg2jwsZzm8CLhDPbiyP7xDIg09jkgIPKVFC71Lyts80VB/vBSbKz0rtZy7a64fu2HU1jpgdRG6PR2hu1nVAD37Zxw965uLPClY6ju5oYe8gsAcvHCoNzyuHAs827jwuzKlw7vvnra8bUlbvCkVe7vdrZu8nxzYvCwbyLyAw7i8dRyRvF3SubsuoeA8geZ1vFrFEzxW8Qe9fu0SvIpWKDjK6P87oHoXPPZLm7xidUo8I6dvvLHj8jwws+a8vSbjvH2H5DiF7Re9IJ7zPKC4M7w/Qxm60BeAvADYa7u4eyo86adXvIax17yzuCq77ZGRu37hibwtzIq8AEspu7o95jwvdW+70t+mvJ6K8jtmBmM79UdqPKRISLsPw9O7HSuiPMp5Aj2FiVy9uArzupmz8jvYcPG7Q5uxvI3PTjzQ4he8EUNbPHsqDDyWaZ48KAlGPUYuz7xEy6y7M7qkPDXcqjzJw8s8Xh+8PAECIbvgKdo6C9BFPWqu4TuWDKa8x5BZu1SLLj0lcik8/4x8vANHGDxppbQ8XQ0Nu7WP9DqOP++7lo2PvHvXsrwmtAi8XC8tvQYJgbw0llA8KwKrO3m8yTxLAni8lbFcvEQNtTzF00W8qDabPNWNZzzCqam8KqgMulHOVDzU81I8HxkzO+bsLT32cy+8ONKQvEUzqDxTct87r7p9PGNvMLu7SqM87apSPDQmq7x9X6S70myaPB654bswXiK9bMs/PA2KqDuEFVU8YPzCPLaPjjyW0Mg8aclxPEZoPDym1p88WaAWPWw66ro1zoQ6LjK4PJQI5rwIPZM8BKKvu8bF9TzW6SO8svqYvLGZ4LwBmyO7w0jIvItfJb0uZTc8iafnu+d31rqKIta6MnCRPAO4HLuRv4a8XmQ7vW/8JL0LVmw77S+GPOpujDzmX1+8oZfPO5e1ULz2RMy7cov+u5nS67tBOwa8hmXiPPhrHjwm8eO83uQBPf250TtS7Ia9NQaNPFLV5rvzyUa8TaOQPPlyQjxQn4A7mqAaPXc2Q73rmYG6CVdEPQar7rtsxOu8m38LPDwP07u/UNY3H9fmvP5lAD1KsOI8111ku9nWFrw9ag29L+8FPeTbIj2gxE+8fNjDvLLt7rsMiSG9E3r2PJeASzuLsFS8ScXIO4KfJL0A5cU8RzLivIjmar29oeq8D7ppPAbuYjyJjsa6ZnRivHiG8bqJpMO7pdt4vB4Q7DlHtHk8gNfuvNe9pzwWP566KFKmPAO4xzsLb4M85PJOvCF467z00Mo8IMQEPPXB9LzF0h29wFYzvBqtjjvoch69W/XuO9sCPzxaoZK7Q4NgOOPZbbz77uC8IjwpPF1pHDwASR08m9IWuu2gWzpzutI8/kMdPC5vE7x92Nk6IecjvIgQ8jzlO0m8Ekt+OznV7TwPJ/+8UNU3PHQexzzcQx69adEEuxDNgLvQfSm93cvIu9Ug6DxrG0u87W0NvUdTCL3f1eW8TGwwPKHOlLwO7ay74jEQO7eKBD2lzvC7W6zSvCgMZbz5w0W8Rxr1On7a8DyE4ae86eqtvC7XqrtpV2W7OuyVPJphSL0axqm7KzOaPLgOq7oeq0U7HGPGPBH5v7on8tI8qrB8uxdti7yawcu8xfmePIacbjw6+jm8pX7Tu5rXszw8B9I8C22mu9XFmDzQWPS7v33avI01iTpLQbe8KkgcvKjWCDxw2GI8sObOPCNgLLzKnoA7GbmUPPXYcbxQl+M6KEaVvOetcbyKywU95K09PNmFgLzAD/o7FGX7OxUmYbzmCfS7/RgvPDgnZ7wURy883oL7vNviozxfz4q88GoxvDdVLDxsYgE7POdQPGzICLwkX5i8tTXqvKm2GT1LYxI8cC83PMLMHD2J2zi7p1UOPZioCj1bY9A8IbQCPQ2JkrwmblO71i2SPMRQgrxayMo80EWZvENuprvlU8i7M81GvOfXgrzQvt28V2N8vNELljv9cd27nVjau1fpmbwmFg68edKuvCZ2l7uUtua8yhWvOsgDcTytWdu7ZrywvJkRhrxmkak8ZT49vF9OCTwnE528Zg0LvBKoNb2bHbc8nk02Ox2UATy02Sq8Er4bvCCsdjw38t48aCWRPOWYu7v/AWc8Q+SGO3ILAbt60R485aO8PO3EVrxnvUS9IQ25u6586ryR7sI8ydE4vW4S7jusXQm8b8usvK54FTwpP8S88wxDvEFShLzLcdE5DCMFvJhYE73WExg9NkMfvVD6kzxKcFA8db5IvIjsMrzakG67XLzPPPbQQzx6qEG8AwzIu6ANG7oVC5g8gj5cPDv4DzzF6AQ8yZ2CPF6IlbpIy/67630wvD3lYby1my88DIOuPB+iMb2ZLau7CWTGuyv8D7xc1Xw7b1aJvO7ud7wjACC9H+zivEu8yDvdy0U8zJ5WPLfuFLxzDp88cqHXvAP+tbwUMHq8/ikGvIQKUDvlEB68rpaMvJYEtbsLqeI8WQlau6LxULxOzI48CPCNPDpI7Tx55ei7Gk1XvIczA7xcIMa8QdOVvBU+R70I47a8xf42PId9Fb0Z9qS6jwiAPLKX/Dx+I/e8N6YjPCWNmTz26Vk8UqvePIgkCLxiC2A8xfSePGqX0Tt2x5S8mBaAvJr3HT2hxMs8PR2SPIeXpLxOAQg84H70O4VfmTx4Phw8XXDTu7ktHzxrVk48GQ07vCg2kDstgHE5RZtBO3RhuTyopzM8OjowvJkFdTlRZ4C8Lp4lvWkqf7ynUlU8SxXnO5aKnTw4lZy6rQi2vBNkj7zkHji8pO2WvNFuvjzgSo08hCRRvLEnrbwYSpg8auQ9uyHZ1Twnvhg8TAMevJkApDz3uAO87aCTvJYLRDzVxYS8ewiPu61SljvMV567C2olvc83HDx1pqg88YY9vEm5nru05w88fwS3PDhgubsKSym979YLPGXryDvR/947Y0msu6ddEbuqSiO97sPoPOc2IDxJ4GO8nnVFPDSamzx/j3U60WLpvG5V1zmwxqm8clMLvDQo3bs88/68UgmVvO8qPryirv+7xFf8O3N9Hzzc+Z6860pUPB1gW7wb2be85xgvPDl6tjm9XYu8+l2vPNY3mbc/EYs7UrCEPKv8kbxJbm08mNtrOz+OYjx2o0k9UIY7OyF1Gz3CZ3g7psQXu7FqJzsa7Cg8dl35O/Ql4ToO76I8Tpexu2zhSDuZtTQ8xlY3PMyEMz37Zq88bwkYvdJo2ruM5Rk8OGvRPFw1nTu7GN477QQXOkXkhbzIPD+7B40kvYWJ7rvnHqE87AERvMfngTzNJAM8s1gsu7O+4Dy3IiU4hG1kO7oikztrgQG7UBJRPAovk7xWi9e8BDA7vLCQxbpROOk8dtiqvBTxCL09+S27C2T0vGPIZjwpdQW7nMa3vLxY/rzTgb48tf0NvDhNtDtZmye8MacBPXBlBL0gNB29iL90PEJMi7xXVvY5mQO7PMTUr7w/RBk6TyPOPLyCDzzw9wo8XrkKPL4vMjo17VI8ouuqPAdXzbuGrVS8rfjlOyRgSLxOoGW70Um+vMEnezyBfC87B/zTu8bNfrtwzbc7zUWIPPOSET142mc8I4/yO4gR2rxMzE48sFq+vHKP7LxRmxA7X+BuPLZurDvQ8tO7e0qHvFBRTTyT06e8oTmSPMtZOjxRr1w7DQtMO8HvqbyisnY88Ik3uz6jj7sc/XU7zAhCvPDNmDxI2mG8oUSBPIqkGLxL8rw8EcK6OxOzvTtoFEu7KhSXPFHBKTrpe8i8lTcPOrn6+jsvMTi88Q2kvGAnDT2ey5G8PWdAOk1y8LtTAY68bfb5vB151TsgeZG8ipjvPAE+EDqy9488LYNvvDhgJ7z6prY6a+dWPCldzzxKq5e8WayVO7s1jDyzlwC8XSAtu1JGOjz8NzA8GECmPCDTt7xLrLu5G/RLPOqmYLwLD7y78bGOO0bbszwakHs8zhkpPLqvurxawCq88CnWu28s4zvl9Ds8pqw2vJLFUbqBjMS8JXNRPLa1EDwCUCs6P+IlPGYqtbxLqU08kPYEOloxjbtPlxQ8nk5MvH8dzbsRWqm8zY0pPBhBh7uCdBw7m6R4uw== + index: 0 + object: embedding + model: qwen3-embedding:4b + object: list + usage: + prompt_tokens: 3 + total_tokens: 3 + status: + code: 200 + message: OK +version: 1 diff --git a/tests/cassettes/test_sandbox/TestDockerSandboxHaikuRAG.test_get_document.yaml b/tests/cassettes/test_sandbox/TestSandboxHaikuRAG.test_get_document.yaml similarity index 100% rename from tests/cassettes/test_sandbox/TestDockerSandboxHaikuRAG.test_get_document.yaml rename to tests/cassettes/test_sandbox/TestSandboxHaikuRAG.test_get_document.yaml diff --git a/tests/cassettes/test_sandbox/TestDockerSandboxHaikuRAG.test_list_documents_with_data.yaml b/tests/cassettes/test_sandbox/TestSandboxHaikuRAG.test_list_documents_with_data.yaml similarity index 100% rename from tests/cassettes/test_sandbox/TestDockerSandboxHaikuRAG.test_list_documents_with_data.yaml rename to tests/cassettes/test_sandbox/TestSandboxHaikuRAG.test_list_documents_with_data.yaml diff --git a/tests/cassettes/test_sandbox/TestDockerSandboxHaikuRAG.test_search_with_data.yaml b/tests/cassettes/test_sandbox/TestSandboxHaikuRAG.test_search_with_data.yaml similarity index 50% rename from tests/cassettes/test_sandbox/TestDockerSandboxHaikuRAG.test_search_with_data.yaml rename to tests/cassettes/test_sandbox/TestSandboxHaikuRAG.test_search_with_data.yaml index b12b11cd..4ca2fa50 100644 --- a/tests/cassettes/test_sandbox/TestDockerSandboxHaikuRAG.test_search_with_data.yaml +++ b/tests/cassettes/test_sandbox/TestSandboxHaikuRAG.test_search_with_data.yaml @@ -39,4 +39,44 @@ interactions: status: code: 200 message: OK +- request: + headers: + accept: + - application/json + accept-encoding: + - gzip, deflate, zstd + connection: + - keep-alive + content-length: + - '73' + content-type: + - application/json + host: + - localhost:11434 + method: POST + parsed_body: + encoding_format: base64 + input: + - fox + model: qwen3-embedding:4b + uri: http://localhost:11434/v1/embeddings + response: + headers: + content-type: + - application/json + transfer-encoding: + - chunked + parsed_body: + data: + - embedding: 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 + index: 0 + object: embedding + model: qwen3-embedding:4b + object: list + usage: + prompt_tokens: 2 + total_tokens: 2 + status: + code: 200 + message: OK version: 1 diff --git a/tests/cassettes/test_sandbox/TestSandboxLLM.test_llm_function.yaml b/tests/cassettes/test_sandbox/TestSandboxLLM.test_llm_function.yaml new file mode 100644 index 00000000..f3dae05e --- /dev/null +++ b/tests/cassettes/test_sandbox/TestSandboxLLM.test_llm_function.yaml @@ -0,0 +1,51 @@ +interactions: +- request: + headers: + accept: + - application/json + accept-encoding: + - gzip, deflate, zstd + connection: + - keep-alive + content-length: + - '143' + content-type: + - application/json + host: + - localhost:11434 + method: POST + parsed_body: + messages: + - content: What is 2 + 2? Reply with just the number. + role: user + model: gpt-oss + reasoning_effort: low + stream: false + uri: http://localhost:11434/v1/chat/completions + response: + headers: + content-length: + - '311' + content-type: + - application/json + parsed_body: + choices: + - finish_reason: stop + index: 0 + message: + content: '4' + reasoning: Just reply 4. + role: assistant + created: 1771924616 + id: chatcmpl-525 + model: gpt-oss + object: chat.completion + system_fingerprint: fp_ollama + usage: + completion_tokens: 16 + prompt_tokens: 81 + total_tokens: 97 + status: + code: 200 + message: OK +version: 1 diff --git a/uv.lock b/uv.lock index 54af3e15..8de6854d 100644 --- a/uv.lock +++ b/uv.lock @@ -180,7 +180,7 @@ wheels = [ [[package]] name = "anthropic" -version = "0.79.0" +version = "0.83.0" source = { registry = "https://pypi.org/simple" } dependencies = [ { name = "anyio" }, @@ -192,9 +192,9 @@ dependencies = [ { name = "sniffio" }, { name = "typing-extensions" }, ] -sdist = { url = "https://files.pythonhosted.org/packages/15/b1/91aea3f8fd180d01d133d931a167a78a3737b3fd39ccef2ae8d6619c24fd/anthropic-0.79.0.tar.gz", hash = "sha256:8707aafb3b1176ed6c13e2b1c9fb3efddce90d17aee5d8b83a86c70dcdcca871", size = 509825, upload-time = "2026-02-07T18:06:18.388Z" } +sdist = { url = "https://files.pythonhosted.org/packages/db/e5/02cd2919ec327b24234abb73082e6ab84c451182cc3cc60681af700f4c63/anthropic-0.83.0.tar.gz", hash = "sha256:a8732c68b41869266c3034541a31a29d8be0f8cd0a714f9edce3128b351eceb4", size = 534058, upload-time = "2026-02-19T19:26:38.904Z" } wheels = [ - { url = "https://files.pythonhosted.org/packages/95/b2/cc0b8e874a18d7da50b0fda8c99e4ac123f23bf47b471827c5f6f3e4a767/anthropic-0.79.0-py3-none-any.whl", hash = "sha256:04cbd473b6bbda4ca2e41dd670fe2f829a911530f01697d0a1e37321eb75f3cf", size = 405918, upload-time = "2026-02-07T18:06:20.246Z" }, + { url = "https://files.pythonhosted.org/packages/5f/75/b9d58e4e2a4b1fc3e75ffbab978f999baf8b7c4ba9f96e60edb918ba386b/anthropic-0.83.0-py3-none-any.whl", hash = "sha256:f069ef508c73b8f9152e8850830d92bd5ef185645dbacf234bb213344a274810", size = 456991, upload-time = "2026-02-19T19:26:40.114Z" }, ] [[package]] @@ -1092,7 +1092,7 @@ name = "ffmpeg-python" version = "0.2.0" source = { registry = "https://pypi.org/simple" } dependencies = [ - { name = "future" }, + { name = "future", marker = "python_full_version < '3.14'" }, ] sdist = { url = "https://files.pythonhosted.org/packages/dd/5e/d5f9105d59c1325759d838af4e973695081fbbc97182baf73afc78dec266/ffmpeg-python-0.2.0.tar.gz", hash = "sha256:65225db34627c578ef0e11c8b1eb528bb35e024752f6f10b78c011f6f64c4127", size = 21543, upload-time = "2019-07-06T00:19:08.989Z" } wheels = [ @@ -1306,30 +1306,6 @@ wheels = [ { url = "https://files.pythonhosted.org/packages/c4/ab/09169d5a4612a5f92490806649ac8d41e3ec9129c636754575b3553f4ea4/googleapis_common_protos-1.72.0-py3-none-any.whl", hash = "sha256:4299c5a82d5ae1a9702ada957347726b167f9f8d1fc352477702a1e851ff4038", size = 297515, upload-time = "2025-11-06T18:29:13.14Z" }, ] -[[package]] -name = "griffe" -version = "2.0.0" -source = { registry = "https://pypi.org/simple" } -dependencies = [ - { name = "griffecli" }, - { name = "griffelib" }, -] -wheels = [ - { url = "https://files.pythonhosted.org/packages/8b/94/ee21d41e7eb4f823b94603b9d40f86d3c7fde80eacc2c3c71845476dddaa/griffe-2.0.0-py3-none-any.whl", hash = "sha256:5418081135a391c3e6e757a7f3f156f1a1a746cc7b4023868ff7d5e2f9a980aa", size = 5214, upload-time = "2026-02-09T19:09:44.105Z" }, -] - -[[package]] -name = "griffecli" -version = "2.0.0" -source = { registry = "https://pypi.org/simple" } -dependencies = [ - { name = "colorama" }, - { name = "griffelib" }, -] -wheels = [ - { url = "https://files.pythonhosted.org/packages/e6/ed/d93f7a447bbf7a935d8868e9617cbe1cadf9ee9ee6bd275d3040fbf93d60/griffecli-2.0.0-py3-none-any.whl", hash = "sha256:9f7cd9ee9b21d55e91689358978d2385ae65c22f307a63fb3269acf3f21e643d", size = 9345, upload-time = "2026-02-09T19:09:42.554Z" }, -] - [[package]] name = "griffelib" version = "2.0.0" @@ -1456,6 +1432,7 @@ dependencies = [ { name = "pathspec" }, { name = "pydantic" }, { name = "pydantic-ai-slim", extra = ["ag-ui", "fastmcp", "logfire", "openai"] }, + { name = "pydantic-monty" }, { name = "python-dotenv" }, { name = "pyyaml" }, { name = "rich" }, @@ -1529,6 +1506,7 @@ requires-dist = [ { name = "pydantic-ai-slim", extras = ["openai", "fastmcp", "logfire", "ag-ui"], specifier = ">=1.46.0" }, { name = "pydantic-ai-slim", extras = ["vertexai"], marker = "extra == 'vertexai'" }, { name = "pydantic-ai-slim", extras = ["voyageai"], marker = "extra == 'voyageai'" }, + { name = "pydantic-monty", specifier = ">=0.0.7" }, { name = "python-dotenv", specifier = ">=1.2.1" }, { name = "pyyaml", specifier = ">=6.0.3" }, { name = "rich", specifier = ">=14.2.0" }, @@ -1976,14 +1954,14 @@ name = "langchain-core" version = "1.2.13" source = { registry = "https://pypi.org/simple" } dependencies = [ - { name = "jsonpatch" }, - { name = "langsmith" }, - { name = "packaging" }, - { name = "pydantic" }, - { name = "pyyaml" }, - { name = "tenacity" }, - { name = "typing-extensions" }, - { name = "uuid-utils" }, + { name = "jsonpatch", marker = "python_full_version < '3.14'" }, + { name = "langsmith", marker = "python_full_version < '3.14'" }, + { name = "packaging", marker = "python_full_version < '3.14'" }, + { name = "pydantic", marker = "python_full_version < '3.14'" }, + { name = "pyyaml", marker = "python_full_version < '3.14'" }, + { name = "tenacity", marker = "python_full_version < '3.14'" }, + { name = "typing-extensions", marker = "python_full_version < '3.14'" }, + { name = "uuid-utils", marker = "python_full_version < '3.14'" }, ] sdist = { url = "https://files.pythonhosted.org/packages/fb/bb/c501ca60556c11ac80d1454bdcac63cb33583ce4e64fc4535ad5a7d5c6ba/langchain_core-1.2.13.tar.gz", hash = "sha256:d2773d0d0130a356378db9a858cfeef64c3d64bc03722f1d4d6c40eb46fdf01b", size = 831612, upload-time = "2026-02-15T07:45:57.014Z" } wheels = [ @@ -1995,7 +1973,7 @@ name = "langchain-text-splitters" version = "1.1.0" source = { registry = "https://pypi.org/simple" } dependencies = [ - { name = "langchain-core" }, + { name = "langchain-core", marker = "python_full_version < '3.14'" }, ] sdist = { url = "https://files.pythonhosted.org/packages/41/42/c178dcdc157b473330eb7cc30883ea69b8ec60078c7b85e2d521054c4831/langchain_text_splitters-1.1.0.tar.gz", hash = "sha256:75e58acb7585dc9508f3cd9d9809cb14751283226c2d6e21fb3a9ae57582ca22", size = 272230, upload-time = "2025-12-14T01:15:38.659Z" } wheels = [ @@ -2007,15 +1985,15 @@ name = "langsmith" version = "0.7.3" source = { registry = "https://pypi.org/simple" } dependencies = [ - { name = "httpx" }, - { name = "orjson", marker = "platform_python_implementation != 'PyPy'" }, - { name = "packaging" }, - { name = "pydantic" }, - { name = "requests" }, - { name = "requests-toolbelt" }, - { name = "uuid-utils" }, - { name = "xxhash" }, - { name = "zstandard" }, + { name = "httpx", marker = "python_full_version < '3.14'" }, + { name = "orjson", marker = "python_full_version < '3.14' and platform_python_implementation != 'PyPy'" }, + { name = "packaging", marker = "python_full_version < '3.14'" }, + { name = "pydantic", marker = "python_full_version < '3.14'" }, + { name = "requests", marker = "python_full_version < '3.14'" }, + { name = "requests-toolbelt", marker = "python_full_version < '3.14'" }, + { name = "uuid-utils", marker = "python_full_version < '3.14'" }, + { name = "xxhash", marker = "python_full_version < '3.14'" }, + { name = "zstandard", marker = "python_full_version < '3.14'" }, ] sdist = { url = "https://files.pythonhosted.org/packages/8d/bc/8172fefad4f2da888a6d564a27d1fb7d4dbf3c640899c2b40c46235cbe98/langsmith-0.7.3.tar.gz", hash = "sha256:0223b97021af62d2cf53c8a378a27bd22e90a7327e45b353e0069ae60d5d6f9e", size = 988575, upload-time = "2026-02-13T23:25:32.916Z" } wheels = [ @@ -3624,20 +3602,20 @@ email = [ [[package]] name = "pydantic-ai-slim" -version = "1.60.0" +version = "1.63.0" source = { registry = "https://pypi.org/simple" } dependencies = [ { name = "genai-prices" }, - { name = "griffe" }, + { name = "griffelib" }, { name = "httpx" }, { name = "opentelemetry-api" }, { name = "pydantic" }, { name = "pydantic-graph" }, { name = "typing-inspection" }, ] -sdist = { url = "https://files.pythonhosted.org/packages/16/97/f73f439f3d415d43f38250f76852121188d0ef6114ec702e84e7d69c301d/pydantic_ai_slim-1.60.0.tar.gz", hash = "sha256:12ba3e6ef933fcb9fc6a307dbdaa43ca15bbc1b8ec77521afd1b7a526d12330f", size = 418839, upload-time = "2026-02-17T00:33:29.672Z" } +sdist = { url = "https://files.pythonhosted.org/packages/da/6d/2b5c0c60b42e6af49830f6a09b5d38fecdb1f20d9659152691eba95613b4/pydantic_ai_slim-1.63.0.tar.gz", hash = "sha256:9377afecdfe4bc17f5c9ed72c758e460703ac5876931aa2f18ace8ac0e69312a", size = 426862, upload-time = "2026-02-23T17:56:36.215Z" } wheels = [ - { url = "https://files.pythonhosted.org/packages/e3/40/8cb494a4d2ba62b5f92ae8bc79a2abbbf8509cb692edd4bc841695187780/pydantic_ai_slim-1.60.0-py3-none-any.whl", hash = "sha256:6865188a225a2979c82bb022a299d438d805d258c6d3f9810f7fe4e3c86af80a", size = 546410, upload-time = "2026-02-17T00:33:21.901Z" }, + { url = "https://files.pythonhosted.org/packages/f2/ca/c4e39eec1cff5a294b64313a8a959b38d326819e0f0a41f48e61ce019a22/pydantic_ai_slim-1.63.0-py3-none-any.whl", hash = "sha256:ed393b0f871b748171f65bec5191c3025b5abb8a4fc616afee17eb9dc2dfa15d", size = 554190, upload-time = "2026-02-23T17:56:29.533Z" }, ] [package.optional-dependencies] @@ -3681,7 +3659,7 @@ vertexai = [ { name = "requests" }, ] voyageai = [ - { name = "voyageai" }, + { name = "voyageai", marker = "python_full_version < '3.14'" }, ] [[package]] @@ -3757,7 +3735,7 @@ wheels = [ [[package]] name = "pydantic-evals" -version = "1.60.0" +version = "1.63.0" source = { registry = "https://pypi.org/simple" } dependencies = [ { name = "anyio" }, @@ -3767,14 +3745,14 @@ dependencies = [ { name = "pyyaml" }, { name = "rich" }, ] -sdist = { url = "https://files.pythonhosted.org/packages/ec/2c/bed606a726b09adc9ee414bdb919ffe499edf7f8c631ba03b0ff3aa34435/pydantic_evals-1.60.0.tar.gz", hash = "sha256:ae3edd6667075acd8ef04c0d6fffb1ebe72c37ff077295fdbd6319e59284580b", size = 54214, upload-time = "2026-02-17T00:33:31.697Z" } +sdist = { url = "https://files.pythonhosted.org/packages/99/43/21b6ddf65b56f7401c344f98e4e6258a02d2868c8a52a8b79c0e0e701029/pydantic_evals-1.63.0.tar.gz", hash = "sha256:eed56a7192e07c8be8cf16e53bb2ef652b4f7f7b8527650ac45fde865a4ecf9d", size = 56365, upload-time = "2026-02-23T17:56:37.71Z" } wheels = [ - { url = "https://files.pythonhosted.org/packages/fe/a2/60790b2c971f6ce78fea2db2b65800489982c341ad9aa6db750070a93dcd/pydantic_evals-1.60.0-py3-none-any.whl", hash = "sha256:7a7414535002cae63ba0d0d9b15c6e72c28252cf120290b18372c9852c91fcfe", size = 65278, upload-time = "2026-02-17T00:33:23.411Z" }, + { url = "https://files.pythonhosted.org/packages/9b/f2/7174ad6abca2457e35a1b902ca4fa78aa8ee72e4ec2e9cd5dc8904014ec9/pydantic_evals-1.63.0-py3-none-any.whl", hash = "sha256:2e92a3af579a5670b2babf2044081d0ef99ab5a9ef141972616d71fd7e5bfd0e", size = 67279, upload-time = "2026-02-23T17:56:31.008Z" }, ] [[package]] name = "pydantic-graph" -version = "1.60.0" +version = "1.63.0" source = { registry = "https://pypi.org/simple" } dependencies = [ { name = "httpx" }, @@ -3782,9 +3760,53 @@ dependencies = [ { name = "pydantic" }, { name = "typing-inspection" }, ] -sdist = { url = "https://files.pythonhosted.org/packages/f9/e6/1cae7cd39ab29f2eebc87c0a82c7ffcdfbe88492d2fb4afaaad91534e1ff/pydantic_graph-1.60.0.tar.gz", hash = "sha256:9710e457c2f8c113fd63629f05174e45bdca917d90c69ec8cf558649f995505f", size = 58492, upload-time = "2026-02-17T00:33:32.66Z" } +sdist = { url = "https://files.pythonhosted.org/packages/7a/c8/aa3cb56552562b799f31e9de291c8bd88306308cfc9647d220dfff2bea18/pydantic_graph-1.63.0.tar.gz", hash = "sha256:5fd98bb22fa6181f0357a6ffad38a3214af12868bd46492d6456c5db434466b4", size = 58528, upload-time = "2026-02-23T17:56:39.118Z" } wheels = [ - { url = "https://files.pythonhosted.org/packages/d8/6c/ce1c0eca77c6efbf7c7168c05a244c2756bde876a52575f750471d522024/pydantic_graph-1.60.0-py3-none-any.whl", hash = "sha256:741fa1e48424b0def86079a01100ad0652e75882f0352cd157232b75ace468a5", size = 72345, upload-time = "2026-02-17T00:33:25.077Z" }, + { url = "https://files.pythonhosted.org/packages/a4/1c/8dcae24c824dd2690fbe7375083b369b10ed1ad773e2b9d1122bb6c0fcdc/pydantic_graph-1.63.0-py3-none-any.whl", hash = "sha256:d9b7a387116f358d470c042b07aa08125cadfcfa8c08ef01769746a489aef0d5", size = 72353, upload-time = "2026-02-23T17:56:32.304Z" }, +] + +[[package]] +name = "pydantic-monty" +version = "0.0.7" +source = { registry = "https://pypi.org/simple" } +sdist = { url = "https://files.pythonhosted.org/packages/f5/e3/0d8b2b025628477c839f894e632f5197872b19df0a86b2ec30fac3b5960a/pydantic_monty-0.0.7.tar.gz", hash = "sha256:2189ea1d7aadab2f95374733d692f51d1206379a4fc7ce18ab46895512e88f92", size = 684705, upload-time = "2026-02-19T14:12:47.235Z" } +wheels = [ + { url = "https://files.pythonhosted.org/packages/ce/7e/ca0884108c3237bb15bb2a1b3f24ddd957b9c750f1ed3211801497941999/pydantic_monty-0.0.7-cp312-cp312-macosx_10_12_x86_64.whl", hash = "sha256:f35e18f284524d26d5f27084e2b93eb40139055bbf0cab6221a043eb5e9ce2dc", size = 6264252, upload-time = "2026-02-19T14:13:59.317Z" }, + { url = "https://files.pythonhosted.org/packages/5a/5b/31f70c7792a857bacbdce90b8aae4629c31a9fec35f0116d91a2fb53241b/pydantic_monty-0.0.7-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:8e5d8924c65bb1ced60785a156e28c73f7f79f164b4f090dc26312c3917ffff7", size = 6133285, upload-time = "2026-02-19T14:14:46.577Z" }, + { url = "https://files.pythonhosted.org/packages/42/56/c92216c0427e8a10a01fa98f29252f6fabd8ca80ca193e0fd30fe28e65c1/pydantic_monty-0.0.7-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:6279a468469d5a3b80d94dd0ab6110cd291a1dfbb057fa7d6dbad1f499be855d", size = 6059856, upload-time = "2026-02-19T14:13:02.594Z" }, + { url = "https://files.pythonhosted.org/packages/b5/a7/bc3e67b12d8a9da65f2677d9a48bc1e055a1d853d48572ba4845a64075cf/pydantic_monty-0.0.7-cp312-cp312-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:cb5feb69a5902d059db5dab269f90423b85a22163668422be47ccac8d7f7c44a", size = 6313780, upload-time = "2026-02-19T14:13:46.45Z" }, + { url = "https://files.pythonhosted.org/packages/c8/97/a9b856b17ee1e54892dafbb7ea29305520cae2dcd8aafe82a26a0edbc33c/pydantic_monty-0.0.7-cp312-cp312-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:a2f1913e9729aa6711092ecbfce764df199a4787fe3e23a7ed74c78bf846579e", size = 6856827, upload-time = "2026-02-19T14:13:40.836Z" }, + { url = "https://files.pythonhosted.org/packages/98/57/2d8184b9f5a0b2b3bb47fdad7061c6a182699824efe6de9d8dd19ee68c0a/pydantic_monty-0.0.7-cp312-cp312-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:ec528f9f4194e6298757ad99e25da47f06f49ae2bc176ee26f49da5eb1dd7849", size = 6870737, upload-time = "2026-02-19T14:13:53.602Z" }, + { url = "https://files.pythonhosted.org/packages/6b/b3/fe3d3eff82b41e517739841a492d7a48ea2daf8e7b822299b848b5d4c0aa/pydantic_monty-0.0.7-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:75405b9186a9acfa49cd66aa339b5a2450d733a2d59fae20cc8be45e45204d5f", size = 6611843, upload-time = "2026-02-19T14:13:49.957Z" }, + { url = "https://files.pythonhosted.org/packages/ff/d2/fdd8fe135ea14e30b40adadc896dda6c805596688998c9bcdfd8d85a16cf/pydantic_monty-0.0.7-cp312-cp312-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:1917e42fce4733f92f5f6ad64ae4d0e87abbf9fb284ef52589ac3e292e928bfe", size = 6692856, upload-time = "2026-02-19T14:13:00.635Z" }, + { url = "https://files.pythonhosted.org/packages/0d/a6/fdde6f8d76aa0cff4b53060b63d7f09dbb19da61a967cb4b3dfd972acf1c/pydantic_monty-0.0.7-cp312-cp312-musllinux_1_1_aarch64.whl", hash = "sha256:d72a5d4f3ee7f9d2630b0379e4cfb397e181eb0b16e8c49a03f80ba6471edb89", size = 6236587, upload-time = "2026-02-19T14:14:32.602Z" }, + { url = "https://files.pythonhosted.org/packages/1b/7e/0580bbc001a39252b2f7da4b7504ac10572e4ca0ec967aebc5a9d752b6f7/pydantic_monty-0.0.7-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:20028220981516912f130986354ef6c926b98778146ca349560cb852e44d9ca6", size = 6672260, upload-time = "2026-02-19T14:13:26.527Z" }, + { url = "https://files.pythonhosted.org/packages/d8/53/578a7b781a5714db5c4b1989c6e876d30caa0adf8a5a4caad89abc306667/pydantic_monty-0.0.7-cp312-cp312-win32.whl", hash = "sha256:e28b1c3ed52892f8ac12ee0f2b535402dfe1cb1e5c18128f1cb69eb8b66c285a", size = 6131085, upload-time = "2026-02-19T14:14:13.296Z" }, + { url = "https://files.pythonhosted.org/packages/56/98/20bd45fcd472937b1b3438b7587e209e3cfd447c30d02f654b86b44adaad/pydantic_monty-0.0.7-cp312-cp312-win_amd64.whl", hash = "sha256:031dfab63ff9d7acdc641852e0d822603038cef1c27c5060900b9fd51cc853d0", size = 6664431, upload-time = "2026-02-19T14:13:29.968Z" }, + { url = "https://files.pythonhosted.org/packages/ec/fe/d8cb6c30d9d7bcc7d3c8d2c349a227e2a83cd1fbe7182f4941896eb35443/pydantic_monty-0.0.7-cp313-cp313-macosx_10_12_x86_64.whl", hash = "sha256:55d36818f8e35872ed35e395b41df8acc460bcdbbfd471fe0c39e293a1d50db5", size = 6262596, upload-time = "2026-02-19T14:14:15.2Z" }, + { url = "https://files.pythonhosted.org/packages/81/7a/f6b4881ca9779bd87eb8d8c0823133b56c232cf09d765c07a7f91d641490/pydantic_monty-0.0.7-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:acb437458d93d54a9658656545fb6c9b396dbe66f68633b3c57bfd2f4aa1d400", size = 6133793, upload-time = "2026-02-19T14:14:06.498Z" }, + { url = "https://files.pythonhosted.org/packages/39/b9/dfcffd95ff233b8c98db9254242d9c10190989762016d18509aa04d43b1b/pydantic_monty-0.0.7-cp313-cp313-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:b196345ffa1997041cb870ea693148feafe270575e2c2963532eedde0e84dedf", size = 6059400, upload-time = "2026-02-19T14:13:38.953Z" }, + { url = "https://files.pythonhosted.org/packages/6b/2e/d6ecef842024267ddf4128613342b8985a7444e74f3a4a312713c913a91a/pydantic_monty-0.0.7-cp313-cp313-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:d471e3cfe77d62edaf43b7f0962b95270ee4243abea12cb8e8cf1ad972dc3612", size = 6312625, upload-time = "2026-02-19T14:13:36.901Z" }, + { url = "https://files.pythonhosted.org/packages/75/2e/e4a2a9fbc3640bcee15b80c2f8ba0f97bf989c58c01d6da187524f71d12b/pydantic_monty-0.0.7-cp313-cp313-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:3a67771afd385579bf3f894ce933fb9e467fba9a632ccf27246e271d448f6f5f", size = 6859902, upload-time = "2026-02-19T14:14:26.67Z" }, + { url = "https://files.pythonhosted.org/packages/d2/4a/7aaf5c793f52e3403892a2de1f5dd18ae38234d82111cc9b7d92443e5b0d/pydantic_monty-0.0.7-cp313-cp313-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:bf662e1bbee4ddd318d5b8bfa9233173045029be0f67f15f816955b246bb7ec0", size = 6870524, upload-time = "2026-02-19T14:13:28.208Z" }, + { url = "https://files.pythonhosted.org/packages/c3/a9/c16f078864a273460923f1371b769c2719e1ce1ad86bc9031e3ed7fb3eae/pydantic_monty-0.0.7-cp313-cp313-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:0151ff59a8a0d9e29ddb448affa33943108121e6e324795646a2f5facaf1a5d8", size = 6611960, upload-time = "2026-02-19T14:14:38.812Z" }, + { url = "https://files.pythonhosted.org/packages/20/d3/b3ef3432558a8cc9551d8b80a028a0f51cd2a518275932e03359eac3dc39/pydantic_monty-0.0.7-cp313-cp313-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:45f17a65134d3a0031e1f54d143770699f6a0ce92a1e74f0ae4914e52370f058", size = 6691834, upload-time = "2026-02-19T14:13:34.661Z" }, + { url = "https://files.pythonhosted.org/packages/ef/56/1ab5d1cbc0edfb522f0c28c9f5a7fc74eea6355234f73833087524d034bf/pydantic_monty-0.0.7-cp313-cp313-musllinux_1_1_aarch64.whl", hash = "sha256:ede6e68cb8a1f7216e26b0b2fb6cd0eae7a92104be8a49d1042e9e428de9262b", size = 6235704, upload-time = "2026-02-19T14:13:48.276Z" }, + { url = "https://files.pythonhosted.org/packages/c4/d1/cdebae67b0543f696ed7daff8587dc8a458e6552b52d5877cb8e55be74b4/pydantic_monty-0.0.7-cp313-cp313-musllinux_1_1_x86_64.whl", hash = "sha256:4b2fd51eea05a0cc37bb91f326efdb1acbcd4b8262dac1c55aeb208e51254978", size = 6671530, upload-time = "2026-02-19T14:12:56.96Z" }, + { url = "https://files.pythonhosted.org/packages/9f/53/c0dacaec260b71050fd6b31d09570f9d74bbb2a2e9586032694e92b9fa59/pydantic_monty-0.0.7-cp313-cp313-win32.whl", hash = "sha256:40f2092970c5899ac2a2784d712a4c7e194b33cd0133315254e4baf141cd6c93", size = 6130341, upload-time = "2026-02-19T14:14:09.701Z" }, + { url = "https://files.pythonhosted.org/packages/3e/05/31490a7a899d8bbb2e513630ea6f591ceb8a111c91fe7573a96c9f6b6327/pydantic_monty-0.0.7-cp313-cp313-win_amd64.whl", hash = "sha256:42cee2646415bb9bd7da428d169783203618a418f960c6f75c7a74d6946d6b31", size = 6664341, upload-time = "2026-02-19T14:14:19.008Z" }, + { url = "https://files.pythonhosted.org/packages/99/15/64aff358df0b822dd22f212fec501e3944edffe978a4ab05530ea641dc68/pydantic_monty-0.0.7-cp314-cp314-macosx_10_12_x86_64.whl", hash = "sha256:c97b2e1dcd0126417892595c1da724a8c4348f7dcec26ba774117bd51bde46f8", size = 6266090, upload-time = "2026-02-19T14:13:12.364Z" }, + { url = "https://files.pythonhosted.org/packages/64/90/7b5a4292eb9993eb8be9d958b5a57764818eeda471e3e79be7da4e9b49ba/pydantic_monty-0.0.7-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:6f40e6e133b309ba733874f5980ab6cf867ec8cea2a6a389a641819cd8dcb7cd", size = 6152219, upload-time = "2026-02-19T14:13:20.734Z" }, + { url = "https://files.pythonhosted.org/packages/43/e4/2740af0157eb3c6f10c16b0d8376b8c9cf0b910720fe90229885bccb4a91/pydantic_monty-0.0.7-cp314-cp314-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:57f2a327b6aa7402a2b2c3ddb3964bd45f12597f8f950dd4c5905b843b353b73", size = 6060942, upload-time = "2026-02-19T14:12:45.601Z" }, + { url = "https://files.pythonhosted.org/packages/f6/26/1cf235c2cc8e219a94ed8b11151280ba89e8020a475b6280c89e62f7275f/pydantic_monty-0.0.7-cp314-cp314-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:771f1f158af0de2480ea2a8862e4c0c7f79e9a13cc3e17529002e1abfc077f95", size = 6315477, upload-time = "2026-02-19T14:14:11.314Z" }, + { url = "https://files.pythonhosted.org/packages/b8/41/0faca7b9d8868822b7177ae941f193f397479bb114d3a6396466167a3198/pydantic_monty-0.0.7-cp314-cp314-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:a4182312ee8c26d8834e76375b2b5c766cb5d86d1dcf1515aa95e02704fcad83", size = 6862130, upload-time = "2026-02-19T14:13:22.375Z" }, + { url = "https://files.pythonhosted.org/packages/34/7b/0f2bd4105a285f50f17af721e83a76ebc1186a9f07a2a29d6d576490a232/pydantic_monty-0.0.7-cp314-cp314-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:8751696fe66fb1bdd429d98fde3a7f4b7dce9cb45f22095e28b96b690a422a2c", size = 6872292, upload-time = "2026-02-19T14:14:08.078Z" }, + { url = "https://files.pythonhosted.org/packages/c9/18/4380820d62d348afb1355814ea674788e28db7a73108a486e0b8898987de/pydantic_monty-0.0.7-cp314-cp314-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:40b73cd1137bb4fd9bf95ed5f87e48d960f7d556c30eeafde6996f908abbf183", size = 6636567, upload-time = "2026-02-19T14:14:01.464Z" }, + { url = "https://files.pythonhosted.org/packages/3e/57/29e5f89a558a6409d514bb2790c72442ec65804cbf1df6a870bbc0038673/pydantic_monty-0.0.7-cp314-cp314-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:c1fa45a66757de5ea45c0809e15643bc2521959dbe2bc231694b22fa189decc9", size = 6693896, upload-time = "2026-02-19T14:14:34.128Z" }, + { url = "https://files.pythonhosted.org/packages/f6/26/5886d0f57ddb5ddf766ee2d0a4b3032267be9efd49a7c95c4d87a0b4b6a9/pydantic_monty-0.0.7-cp314-cp314-musllinux_1_1_aarch64.whl", hash = "sha256:ffe8122db9f0f64619a66f4cee2f577245ba59158b7d651a2eae08df691d34f9", size = 6236867, upload-time = "2026-02-19T14:14:20.704Z" }, + { url = "https://files.pythonhosted.org/packages/b5/72/1bb8741baf84f217d92291b862f8a8cb64d735fe4be20be2827fdf787593/pydantic_monty-0.0.7-cp314-cp314-musllinux_1_1_x86_64.whl", hash = "sha256:2122a5b6df53843329af01f6671300747449608d9cdf88ae5f9cf9977794e7a2", size = 6673504, upload-time = "2026-02-19T14:14:24.517Z" }, + { url = "https://files.pythonhosted.org/packages/d1/14/a4ff2bfe46350ffde4b5edc1f293b252cff90063f1f4cece49affe5a6462/pydantic_monty-0.0.7-cp314-cp314-win32.whl", hash = "sha256:bfbea2eddb9eef186326a6dfb27d79f8de434d7a3979f36f03b04216234a0275", size = 6131872, upload-time = "2026-02-19T14:13:14.437Z" }, + { url = "https://files.pythonhosted.org/packages/60/1f/d873f280aae5cbd27021189843fbd5f77be4262a7654ef30445d984518ab/pydantic_monty-0.0.7-cp314-cp314-win_amd64.whl", hash = "sha256:1b750afceef78f5c5d3e3e3c32a8060b3a7e1b97e3a00ac2bede5bc5e87cde8f", size = 6687125, upload-time = "2026-02-19T14:13:04.317Z" }, ] [[package]] @@ -4352,7 +4374,7 @@ name = "requests-toolbelt" version = "1.0.0" source = { registry = "https://pypi.org/simple" } dependencies = [ - { name = "requests" }, + { name = "requests", marker = "python_full_version < '3.14'" }, ] sdist = { url = "https://files.pythonhosted.org/packages/f3/61/d7545dafb7ac2230c70d38d31cbfe4cc64f7144dc41f6e4e4b78ecd9f5bb/requests-toolbelt-1.0.0.tar.gz", hash = "sha256:7681a0a3d047012b5bdc0ee37d7f8f07ebe76ab08caeccfc3921ce23c88d5bc6", size = 206888, upload-time = "2023-05-01T04:11:33.229Z" } wheels = [ @@ -5325,16 +5347,16 @@ name = "voyageai" version = "0.3.7" source = { registry = "https://pypi.org/simple" } dependencies = [ - { name = "aiohttp" }, - { name = "aiolimiter" }, - { name = "ffmpeg-python" }, - { name = "langchain-text-splitters" }, + { name = "aiohttp", marker = "python_full_version < '3.14'" }, + { name = "aiolimiter", marker = "python_full_version < '3.14'" }, + { name = "ffmpeg-python", marker = "python_full_version < '3.14'" }, + { name = "langchain-text-splitters", marker = "python_full_version < '3.14'" }, { name = "numpy", marker = "python_full_version < '3.14'" }, - { name = "pillow" }, - { name = "pydantic" }, - { name = "requests" }, - { name = "tenacity" }, - { name = "tokenizers" }, + { name = "pillow", marker = "python_full_version < '3.14'" }, + { name = "pydantic", marker = "python_full_version < '3.14'" }, + { name = "requests", marker = "python_full_version < '3.14'" }, + { name = "tenacity", marker = "python_full_version < '3.14'" }, + { name = "tokenizers", marker = "python_full_version < '3.14'" }, ] sdist = { url = "https://files.pythonhosted.org/packages/94/16/1b46b3cd401e1717a68197c1fe336d7bb4e0a1833f8105e1738f5b1add05/voyageai-0.3.7.tar.gz", hash = "sha256:826cd97f97223f42b5babc5c459c9c80f3a8215ce5c0e007b0b276550f790d24", size = 26485, upload-time = "2025-12-16T18:43:05.26Z" } wheels = [