Rank q/a history with respect to cosine similarity to new queries. Adapt the chat agent to only pass relevant q/as

haiku_rag_slim/haiku/rag/agents/chat/agent.py

@@ -10,6 +10,7 @@ from haiku.rag.agents.chat.state import (
     QAResponse,
     build_document_filter,
     format_conversation_context,
+    rank_qa_history_by_similarity,
 )
 from haiku.rag.agents.research.dependencies import ResearchContext
 from haiku.rag.agents.research.graph import build_conversational_graph
@@ -133,10 +134,21 @@ def create_chat_agent(config: AppConfig) -> Agent[ChatDeps, str]:
         # Build filter from document_name
         doc_filter = build_document_filter(document_name) if document_name else None
 
-        # Convert existing qa_history to SearchAnswers for context seeding
-        existing_qa: list[SearchAnswer] = []
+        # Rank qa_history by similarity to current question
+        ranked_history: list[QAResponse] = []
         if ctx.deps.session_state and ctx.deps.session_state.qa_history:
-            for qa in ctx.deps.session_state.qa_history:
+            embedder = ctx.deps.client.chunk_repository.embedder
+            ranked_history = await rank_qa_history_by_similarity(
+                current_question=question,
+                qa_history=ctx.deps.session_state.qa_history,
+                embedder=embedder,
+                top_k=5,
+            )
+
+        # Convert ranked qa_history to SearchAnswers for context seeding
+        existing_qa: list[SearchAnswer] = []
+        if ranked_history:
+            for qa in ranked_history:
                 citations = [
                     Citation(
                         document_id=c.document_id,

haiku_rag_slim/haiku/rag/agents/chat/state.py

@@ -1,5 +1,9 @@
+import logging
 from dataclasses import dataclass, field
+from typing import TYPE_CHECKING
 
+import numpy as np
+from numpy.typing import NDArray
 from pydantic import BaseModel
 from pydantic_ai import format_as_xml
 
@@ -7,6 +11,11 @@ from haiku.rag.client import HaikuRAG
 from haiku.rag.config.models import AppConfig
 from haiku.rag.store.models import SearchResult
 
+if TYPE_CHECKING:
+    from haiku.rag.embeddings import EmbedderWrapper
+
+logger = logging.getLogger(__name__)
+
 
 class CitationInfo(BaseModel):
     """Citation info for frontend display."""
@@ -63,6 +72,65 @@ def format_conversation_context(qa_history: list[QAResponse]) -> str:
     return format_as_xml(context_data, root_tag="conversation_context")
 
 
+def _cosine_similarity(a: NDArray[np.float64], b: NDArray[np.float64]) -> float:
+    """Compute cosine similarity between two vectors."""
+    return float(np.dot(a, b) / (np.linalg.norm(a) * np.linalg.norm(b)))
+
+
+async def rank_qa_history_by_similarity(
+    current_question: str,
+    qa_history: list[QAResponse],
+    embedder: "EmbedderWrapper",
+    top_k: int = 5,
+) -> list[QAResponse]:
+    """Rank Q&A history by semantic similarity to current question.
+
+    Embeds question+answer pairs and returns the top-K most similar to the
+    current question. Falls back to returning the last top_k entries if
+    embedding fails.
+
+    Args:
+        current_question: The current question to compare against.
+        qa_history: List of previous Q&A pairs.
+        embedder: Embedder instance to use for embedding.
+        top_k: Maximum number of entries to return.
+
+    Returns:
+        Top-K Q&A pairs ranked by similarity to current question.
+    """
+    if not qa_history:
+        return []
+
+    if len(qa_history) <= top_k:
+        return qa_history
+
+    try:
+        # Embed current question
+        question_embedding = np.array(await embedder.embed_query(current_question))
+
+        # Embed Q&A pairs as "Q: {question}\nA: {answer}"
+        qa_texts = [f"Q: {qa.question}\nA: {qa.answer}" for qa in qa_history]
+        qa_embeddings = await embedder.embed_documents(qa_texts)
+
+        # Compute similarities
+        similarities: list[tuple[int, float]] = []
+        for i, qa_emb in enumerate(qa_embeddings):
+            sim = _cosine_similarity(question_embedding, np.array(qa_emb))
+            similarities.append((i, sim))
+
+        # Sort by similarity (descending) and take top-K
+        similarities.sort(key=lambda x: x[1], reverse=True)
+        top_indices = sorted([idx for idx, _ in similarities[:top_k]])
+
+        # Return in original order
+        return [qa_history[i] for i in top_indices]
+
+    except Exception as e:
+        logger.warning(f"Failed to rank qa_history by similarity: {e}")
+        # Fallback: return last top_k entries
+        return qa_history[-top_k:]
+
+
 @dataclass
 class ChatDeps:
     """Dependencies for chat agent."""

tests/agents/chat/test_chat_agent.py

@@ -1,3 +1,7 @@
+from pathlib import Path
+
+import pytest
+
 from haiku.rag.agents.chat import (
     ChatDeps,
     ChatSessionState,
@@ -10,6 +14,11 @@ from haiku.rag.client import HaikuRAG
 from haiku.rag.config import Config
 
 
+@pytest.fixture(scope="module")
+def vcr_cassette_dir():
+    return str(Path(__file__).parent.parent.parent / "cassettes" / "test_chat_agent")
+
+
 def test_create_chat_agent():
     """Test that create_chat_agent returns a properly configured agent."""
     agent = create_chat_agent(Config)
@@ -103,3 +112,118 @@ def test_search_agent_initialization(temp_db_path):
     search_agent = SearchAgent(client, Config)
     assert search_agent is not None
     client.close()
+
+
+# DocLayNet content for testing
+DOCLAYNET_CLASS_LABELS = """
+DocLayNet Dataset - Class Labels
+
+DocLayNet defines 11 distinct class labels for document layout analysis:
+1. Caption - Text describing figures or tables
+2. Footnote - Notes at the bottom of pages
+3. Formula - Mathematical expressions
+4. List-item - Items in bulleted or numbered lists
+5. Page-footer - Footer content on pages
+6. Page-header - Header content on pages
+7. Picture - Images and diagrams
+8. Section-header - Headings for document sections
+9. Table - Tabular data
+10. Text - Regular paragraph text (highest count: 510,377 instances)
+11. Title - Document titles
+
+The Text class has the highest count with 510,377 instances in the dataset.
+"""
+
+DOCLAYNET_ANNOTATION = """
+DocLayNet Dataset - Annotation Process
+
+The annotation process was organized into 4 phases:
+- Phase 1: Data selection and preparation by a small team of experts
+- Phase 2: Label selection and guideline definition
+- Phase 3: Annotation by 40 dedicated annotators
+- Phase 4: Quality control and continuous supervision
+
+The Corpus Conversion Service (CCS) was used for annotation, providing a visual interface.
+"""
+
+DOCLAYNET_DATA_SOURCES = """
+DocLayNet Dataset - Data Sources
+
+The data sources for DocLayNet include:
+- Publication repositories such as arXiv
+- Government offices and official documents
+- Company websites and corporate reports
+- Data directory services for financial reports
+- Patent documents
+
+Scanned documents were excluded to avoid rotation and skewing issues.
+"""
+
+
+@pytest.mark.asyncio
+@pytest.mark.vcr()
+async def test_chat_agent_with_qa_history_ranking(allow_model_requests, temp_db_path):
+    """Test chat agent uses similarity ranking for qa_history.
+
+    This test verifies that when qa_history has more than 5 entries,
+    the ranking function is applied and the agent can still process requests.
+    """
+    async with HaikuRAG(temp_db_path, create=True) as client:
+        # Add a simple document
+        await client.create_document(
+            content=DOCLAYNET_CLASS_LABELS,
+            uri="doclaynet-labels",
+            title="DocLayNet Class Labels",
+        )
+
+        agent = create_chat_agent(Config)
+
+        # Build session state with pre-populated qa_history (>5 items to trigger ranking)
+        session_state = ChatSessionState(
+            session_id="test-ranking",
+            qa_history=[
+                QAResponse(
+                    question="What are the 11 class labels in DocLayNet?",
+                    answer="The 11 class labels are: Caption, Footnote, Formula, List-item, Page-footer, Page-header, Picture, Section-header, Table, Text, and Title.",
+                ),
+                QAResponse(
+                    question="How was the annotation process organized?",
+                    answer="The annotation was organized into 4 phases.",
+                ),
+                QAResponse(
+                    question="What data sources were used?",
+                    answer="Sources include arXiv and government offices.",
+                ),
+                QAResponse(
+                    question="How were pages selected?",
+                    answer="By selective subsampling.",
+                ),
+                QAResponse(
+                    question="What is the agreement metric?",
+                    answer="The mAP metric was used.",
+                ),
+                QAResponse(
+                    question="What is machine learning?",
+                    answer="A field of AI.",
+                ),
+            ],
+        )
+
+        deps = ChatDeps(
+            client=client,
+            config=Config,
+            session_state=session_state,
+        )
+
+        # Ask a question - the key test is that ranking is applied without error
+        result = await agent.run(
+            "What class labels are defined in the dataset?",
+            deps=deps,
+        )
+
+        # Verify the agent produced a response (ranking didn't break anything)
+        assert result.output is not None
+        assert len(result.output) > 0
+
+        # Verify qa_history was updated (new Q&A was added)
+        assert len(session_state.qa_history) == 7  # 6 original + 1 new

tests/agents/chat/test_state.py

@@ -0,0 +1,168 @@
+from pathlib import Path
+
+import pytest
+
+from haiku.rag.agents.chat.state import (
+    CitationInfo,
+    QAResponse,
+    rank_qa_history_by_similarity,
+)
+from haiku.rag.client import HaikuRAG
+
+
+@pytest.fixture(scope="module")
+def vcr_cassette_dir():
+    return str(Path(__file__).parent.parent.parent / "cassettes" / "test_chat_state")
+
+
+@pytest.mark.asyncio
+async def test_rank_qa_history_empty():
+    """Test empty history returns empty list."""
+    # Create a mock embedder - we won't actually call it
+    result = await rank_qa_history_by_similarity(
+        current_question="What is this?",
+        qa_history=[],
+        embedder=None,  # type: ignore - won't be called for empty list
+        top_k=5,
+    )
+    assert result == []
+
+
+@pytest.mark.asyncio
+@pytest.mark.vcr()
+async def test_rank_qa_history_small_list(temp_db_path, allow_model_requests):
+    """Test with history smaller than top_k returns all entries."""
+    async with HaikuRAG(temp_db_path, create=True) as client:
+        embedder = client.chunk_repository.embedder
+
+        # Create 3 Q&A pairs (less than top_k=5)
+        qa_history = [
+            QAResponse(question="What is Python?", answer="A programming language"),
+            QAResponse(question="What is Java?", answer="Another programming language"),
+            QAResponse(
+                question="What is Rust?", answer="A systems programming language"
+            ),
+        ]
+
+        result = await rank_qa_history_by_similarity(
+            current_question="Tell me about Python",
+            qa_history=qa_history,
+            embedder=embedder,
+            top_k=5,
+        )
+
+        # Should return all 3 entries since history < top_k
+        assert len(result) == 3
+        # All original entries should be present
+        assert set(qa.question for qa in result) == {
+            "What is Python?",
+            "What is Java?",
+            "What is Rust?",
+        }
+
+
+@pytest.mark.asyncio
+@pytest.mark.vcr()
+async def test_rank_qa_history_returns_top_k(temp_db_path, allow_model_requests):
+    """Test ranking returns top-K most similar entries."""
+    async with HaikuRAG(temp_db_path, create=True) as client:
+        embedder = client.chunk_repository.embedder
+
+        # Create 10 Q&A pairs on different topics
+        qa_history = [
+            QAResponse(
+                question="What are the 11 class labels in DocLayNet?",
+                answer="Caption, Footnote, Formula, List-item, Page-footer, Page-header, Picture, Section-header, Table, Text, and Title",
+            ),
+            QAResponse(
+                question="How was the annotation process organized?",
+                answer="The process had 4 phases with 40 dedicated annotators",
+            ),
+            QAResponse(
+                question="What data sources were used?",
+                answer="arXiv, government offices, company websites, financial reports and patents",
+            ),
+            QAResponse(
+                question="How were pages selected?",
+                answer="By selective subsampling with bias towards pages with figures or tables",
+            ),
+            QAResponse(
+                question="What is the inter-annotator agreement?",
+                answer="Computed as mAP@0.5-0.95 metric between pairwise annotations",
+            ),
+            QAResponse(
+                question="What is machine learning?",
+                answer="A field of AI that enables systems to learn from data",
+            ),
+            QAResponse(
+                question="How does neural network training work?",
+                answer="Through backpropagation and gradient descent",
+            ),
+            QAResponse(
+                question="What is deep learning?",
+                answer="A subset of ML using neural networks with many layers",
+            ),
+        ]
+
+        # Ask a question related to class labels (Q1)
+        result = await rank_qa_history_by_similarity(
+            current_question="Which class label has the highest count in DocLayNet?",
+            qa_history=qa_history,
+            embedder=embedder,
+            top_k=5,
+        )
+
+        # Should return exactly 5 entries
+        assert len(result) == 5
+
+        # The class labels Q&A should be in the top 5 (it's most semantically similar)
+        result_questions = [qa.question for qa in result]
+        assert "What are the 11 class labels in DocLayNet?" in result_questions
+
+
+@pytest.mark.asyncio
+@pytest.mark.vcr()
+async def test_rank_qa_history_preserves_order(temp_db_path, allow_model_requests):
+    """Test that ranking preserves original order among selected items."""
+    async with HaikuRAG(temp_db_path, create=True) as client:
+        embedder = client.chunk_repository.embedder
+
+        # Create Q&A pairs where multiple are similar
+        qa_history = [
+            QAResponse(
+                question="What is Python?",
+                answer="A programming language",
+                citations=[
+                    CitationInfo(
+                        index=1,
+                        document_id="doc1",
+                        chunk_id="chunk1",
+                        document_uri="python.md",
+                        content="Python content",
+                    )
+                ],
+            ),
+            QAResponse(
+                question="What is Java?",
+                answer="Another programming language",
+            ),
+            QAResponse(
+                question="How to use Python for data science?",
+                answer="Use pandas, numpy, and scikit-learn",
+            ),
+        ]
+
+        result = await rank_qa_history_by_similarity(
+            current_question="Tell me about Python programming",
+            qa_history=qa_history,
+            embedder=embedder,
+            top_k=3,
+        )
+
+        # All should be returned
+        assert len(result) == 3
+
+        # The two Python-related questions should be in the results
+        result_questions = [qa.question for qa in result]
+        assert "What is Python?" in result_questions
+        assert "How to use Python for data science?" in result_questions