haiku.rag/docs/agents.md

Agents

Four agentic flows are provided by haiku.rag:

• Simple QA Agent — a focused question answering agent
• Chat Agent — multi-turn conversational RAG with session memory
• Research Graph — a multi-step research workflow with question decomposition
• RLM Agent — complex analytical tasks via sandboxed Python code execution (see RLM Agent)

See QA and Research Configuration for configuring model, iterations, concurrency, and other settings.

Simple QA Agent

The simple QA agent answers a single question using the knowledge base. It retrieves relevant chunks, optionally expands context around them, and asks the model to answer strictly based on that context.

Key points:

• Uses a single search_documents tool to fetch relevant chunks
• Can be run with or without inline citations in the prompt
• Returns a plain string answer

CLI usage:

haiku-rag ask "What is climate change?"

# With citations
haiku-rag ask "What is climate change?" --cite

# Deep mode (uses research graph with optimized settings)
haiku-rag ask "What are the main features of haiku.rag?" --deep

Python usage:

from haiku.rag.client import HaikuRAG
from haiku.rag.agents.qa.agent import QuestionAnswerAgent

async with HaikuRAG(path_to_db) as client:
    agent = QuestionAnswerAgent(
        client=client,
        provider="openai",
        model="gpt-4o-mini",
        use_citations=False,
    )

    answer = await agent.answer("What is climate change?")
    print(answer)

Chat Agent

The chat agent enables multi-turn conversational RAG. It is built from composable toolsets and maintains session state to improve follow-up answers.

Key features:

• Composable toolsets: Built from reusable FunctionToolset factories — see Toolsets
• Semantic prior answer recall: Similar prior Q/A pairs are retrieved and passed to the research planner, which can skip searching when they suffice
• Background summarization: After each ask call, the QA history is summarized into a compact session context for the next request
• Document filtering: Session-level or per-query document filtering

CLI Usage

haiku-rag chat
haiku-rag chat --db /path/to/database.lancedb

See Applications for the full TUI interface guide.

Python Usage

from haiku.rag.client import HaikuRAG
from haiku.rag.agents.chat import create_chat_agent, ChatDeps
from haiku.rag.tools import ToolContext

async with HaikuRAG(path_to_db) as client:
    # Create agent with composed toolsets
    context = ToolContext()
    agent = create_chat_agent(config, client, context)
    deps = ChatDeps(config=config, tool_context=context)

    # First question
    result = await agent.run("What is haiku.rag?", deps=deps)
    print(result.output)

    # Follow-up (uses session context)
    result = await agent.run("How does it handle PDFs?", deps=deps)
    print(result.output)

Feature Selection

By default, create_chat_agent enables search, documents, and QA toolsets. You can customize which capabilities the agent has via the features parameter:

from haiku.rag.agents.chat import (
    create_chat_agent,
    FEATURE_SEARCH,
    FEATURE_DOCUMENTS,
    FEATURE_QA,
    FEATURE_ANALYSIS,
)

# Search-only agent
agent = create_chat_agent(config, client, context, features=[FEATURE_SEARCH])

# All features including code analysis
agent = create_chat_agent(
    config, client, context,
    features=[FEATURE_SEARCH, FEATURE_DOCUMENTS, FEATURE_QA, FEATURE_ANALYSIS],
)

Available features:

Feature	Constant	Tools added
Search	FEATURE_SEARCH	search
Documents	FEATURE_DOCUMENTS	list_documents, get_document, summarize_document
QA	FEATURE_QA	ask
Analysis	FEATURE_ANALYSIS	analyze

The system prompt is automatically composed to match the selected features. See Toolsets for details on each toolset's parameters and behavior.

Session State

The ChatSessionState maintains:

• session_id — Unique identifier for the session
• qa_history — List of previous Q/A pairs
• session_context — Automatically maintained session context summary
• document_filter — List of document titles/URIs to restrict searches
• citation_registry — Stable mapping of chunk IDs to citation indices

Citation Registry: Citation indices persist across tool calls within a session. The same chunk_id always returns the same citation index (first-occurrence-wins). This ensures consistent citation numbering in multi-turn conversations — [1] always refers to the same source.

Conversational Memory

The chat agent maintains two layers of conversational memory:

1. Semantic prior answer recall

When the ask tool receives a question, it embeds the question and compares it against prior Q/A embeddings. Sufficiently similar prior answers are passed to the research planner, which can skip searching entirely if they already cover the question.

2. Background session summarization

After each ask call, a background task summarizes the full QA history into a compact session context. This summary is injected into the research planner on the next request, allowing it to resolve ambiguous references ("Tell me more about the authentication part") without having seen the full conversation.

Research Graph

The research workflow is implemented as a typed pydantic-graph. It uses an iterative feedback loop where the planner proposes one question at a time, sees the answer, then decides whether to continue or synthesize.

---
title: Research graph
---
stateDiagram-v2
  state plan_next_decision <<choice>>
  [*] --> plan_next
  plan_next --> plan_next_decision
  plan_next_decision --> search_one: Has next question
  plan_next_decision --> synthesize: Complete or max iterations
  search_one --> plan_next: Answer added to context
  synthesize --> [*]

  note right of plan_next
    Receives session_context as background
    and prior_answers from conversation history.
    Uses a different prompt when prior answers exist.
  end note

The graph receives a ResearchContext containing:

• original_question — the user's question
• session_context — summary of conversation history (injected as <background> XML)
• qa_responses — prior answers from semantic matching or previous iterations (injected as <prior_answers> XML)

When prior answers are provided, the planner uses a context-aware prompt that evaluates whether existing evidence is sufficient. If it is, the planner marks is_complete=True and the graph skips directly to synthesis without any searches.

Key nodes:

• plan_next: Evaluates gathered evidence and either proposes the next question to investigate or marks research as complete. Uses a context-aware prompt when prior answers exist, allowing it to skip research entirely.
• search_one: Answers a single question using the knowledge base (up to 3 search calls per question). Each answer is added to ResearchContext.qa_responses for the next planning iteration.
• synthesize: Generates the final output from all gathered evidence.

Output modes:

The graph supports two output modes via build_research_graph(output_mode=...):

Mode	Output type	Used by
"report"	ResearchReport (title, executive summary, findings, conclusions, recommendations)	CLI haiku-rag research, Python API
"conversational"	ConversationalAnswer (answer, citations, confidence)	Chat agent's ask tool

Iterative flow:

• Each iteration: planner evaluates context → proposes one question → search answers it → loop back
• Planner can decompose complex questions (e.g., "benefits and drawbacks" → start with "benefits")
• Session context resolves ambiguous references and informs planning
• Prior answers let the planner skip redundant searches
• Loop terminates when planner marks is_complete=True or max_iterations is reached

CLI Usage

# Basic usage
haiku-rag research "How does haiku.rag organize and query documents?"

# With document filter
haiku-rag research "What are the key findings?" --filter "uri LIKE '%report%'"

Python Usage

Basic example:

from haiku.rag.client import HaikuRAG
from haiku.rag.config import Config
from haiku.rag.agents.research.dependencies import ResearchContext
from haiku.rag.agents.research.graph import build_research_graph
from haiku.rag.agents.research.state import ResearchDeps, ResearchState

async with HaikuRAG(path_to_db) as client:
    graph = build_research_graph(config=Config)
    context = ResearchContext(original_question="What are the main features?")
    state = ResearchState.from_config(context=context, config=Config)
    deps = ResearchDeps(client=client)

    report = await graph.run(state=state, deps=deps)

    print(report.title)
    print(report.executive_summary)

With custom config:

from haiku.rag.client import HaikuRAG
from haiku.rag.config.models import AppConfig, ResearchConfig
from haiku.rag.agents.research.dependencies import ResearchContext
from haiku.rag.agents.research.graph import build_research_graph
from haiku.rag.agents.research.state import ResearchDeps, ResearchState

custom_config = AppConfig(
    research=ResearchConfig(
        provider="openai",
        model="gpt-4o-mini",
        max_iterations=5,
        max_concurrency=3,
    )
)

async with HaikuRAG(path_to_db) as client:
    graph = build_research_graph(config=custom_config)
    context = ResearchContext(original_question="What are the main features?")
    state = ResearchState.from_config(context=context, config=custom_config)
    deps = ResearchDeps(client=client)

    report = await graph.run(state=state, deps=deps)

Conversational mode with prior answers:

from haiku.rag.agents.research.dependencies import ResearchContext
from haiku.rag.agents.research.graph import build_research_graph
from haiku.rag.agents.research.models import SearchAnswer
from haiku.rag.agents.research.state import ResearchDeps, ResearchState

# Conversational mode returns ConversationalAnswer instead of ResearchReport
graph = build_research_graph(config=Config, output_mode="conversational")

# Pass session context and prior answers from conversation history
context = ResearchContext(
    original_question="How does it handle authentication?",
    session_context="User is building a Python web app with FastAPI.",
    qa_responses=[
        SearchAnswer(
            query="What authentication methods are supported?",
            answer="JWT and OAuth2 are supported.",
            confidence=0.95,
            cited_chunks=["chunk-1"],
        )
    ],
)
state = ResearchState.from_config(context=context, config=Config)
deps = ResearchDeps(client=client)

result = await graph.run(state=state, deps=deps)
print(result.answer)       # Direct conversational answer
print(result.confidence)   # 0.0-1.0
print(result.citations)    # Deduplicated citations from all searches

Filtering Documents

Restrict searches to specific documents via the search_filter parameter:

# Set filter before running the graph
state = ResearchState.from_config(context=context, config=Config)
state.search_filter = "id IN ('doc-123', 'doc-456')"

report = await graph.run(state=state, deps=deps)

The filter applies to all search operations in the graph. See Filtering Search Results for available filter columns and syntax.