feat: langchain retriever wrapper over existing async hybrid retriever

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

agent_bench/langchain_baseline/__init__.py

@@ -0,0 +1 @@
+"""LangChain baseline: tool-calling agent for framework comparison."""

agent_bench/langchain_baseline/retriever.py

@@ -0,0 +1,66 @@
+"""LangChain BaseRetriever wrapping agent-bench's async hybrid retriever."""
+
+from __future__ import annotations
+
+import asyncio
+from typing import TYPE_CHECKING, Any, List
+
+from langchain_core.callbacks import (
+    AsyncCallbackManagerForRetrieverRun,
+    CallbackManagerForRetrieverRun,
+)
+from langchain_core.documents import Document as LCDocument
+from langchain_core.retrievers import BaseRetriever
+
+if TYPE_CHECKING:
+    from agent_bench.rag.retriever import Retriever
+
+
+class AgentBenchRetriever(BaseRetriever):
+    """Wraps agent-bench's async Retriever as a LangChain retriever.
+
+    Delegates to Retriever.search() which returns list[SearchResult].
+    Each SearchResult has .chunk.content, .chunk.source, .chunk.id, .score.
+    """
+
+    retriever: Any  # agent_bench.rag.retriever.Retriever (Pydantic can't validate it)
+    top_k: int = 5
+
+    model_config = {"arbitrary_types_allowed": True}
+
+    async def _aget_relevant_documents(
+        self,
+        query: str,
+        *,
+        run_manager: AsyncCallbackManagerForRetrieverRun,
+    ) -> List[LCDocument]:
+        results = await self.retriever.search(query, top_k=self.top_k)
+        return [
+            LCDocument(
+                page_content=r.chunk.content,
+                metadata={
+                    "source": r.chunk.source,
+                    "chunk_id": r.chunk.id,
+                    "score": r.score,
+                },
+            )
+            for r in results
+        ]
+
+    def _get_relevant_documents(
+        self,
+        query: str,
+        *,
+        run_manager: CallbackManagerForRetrieverRun,
+    ) -> List[LCDocument]:
+        """Sync fallback: runs async implementation in a new event loop thread."""
+        loop = asyncio.new_event_loop()
+        try:
+            return loop.run_until_complete(
+                self._aget_relevant_documents(
+                    query,
+                    run_manager=AsyncCallbackManagerForRetrieverRun.get_noop_manager(),
+                )
+            )
+        finally:
+            loop.close()

tests/test_langchain_baseline/test_retriever.py

@@ -0,0 +1,71 @@
+"""Tests for LangChain retriever wrapper around agent-bench's async Retriever."""
+
+from unittest.mock import AsyncMock, MagicMock
+
+import pytest
+
+from agent_bench.langchain_baseline.retriever import AgentBenchRetriever
+
+
+def _make_mock_retriever(results=None):
+    """Create a mock of agent_bench.rag.retriever.Retriever."""
+    retriever = MagicMock()
+    if results is None:
+        # Default: one result with known fields
+        result = MagicMock()
+        result.chunk.content = "Path parameters use curly braces."
+        result.chunk.source = "fastapi_path_params.md"
+        result.chunk.id = "chunk_001"
+        result.score = 0.85
+        result.rank = 1
+        results = [result]
+    retriever.search = AsyncMock(return_value=results)
+    return retriever
+
+
+async def test_returns_langchain_documents():
+    mock_ret = _make_mock_retriever()
+    wrapper = AgentBenchRetriever(retriever=mock_ret, top_k=5)
+    docs = await wrapper.ainvoke("path parameters")
+
+    assert len(docs) == 1
+    assert docs[0].page_content == "Path parameters use curly braces."
+    assert docs[0].metadata["source"] == "fastapi_path_params.md"
+    assert docs[0].metadata["chunk_id"] == "chunk_001"
+    assert docs[0].metadata["score"] == 0.85
+
+
+async def test_passes_top_k_to_underlying_retriever():
+    mock_ret = _make_mock_retriever()
+    wrapper = AgentBenchRetriever(retriever=mock_ret, top_k=3)
+    await wrapper.ainvoke("test")
+    mock_ret.search.assert_called_once_with("test", top_k=3)
+
+
+async def test_handles_empty_results():
+    mock_ret = _make_mock_retriever(results=[])
+    wrapper = AgentBenchRetriever(retriever=mock_ret, top_k=5)
+    docs = await wrapper.ainvoke("nonsense")
+    assert docs == []
+
+
+async def test_multiple_results_preserve_order():
+    r1 = MagicMock()
+    r1.chunk.content = "First"
+    r1.chunk.source = "a.md"
+    r1.chunk.id = "c1"
+    r1.score = 0.9
+
+    r2 = MagicMock()
+    r2.chunk.content = "Second"
+    r2.chunk.source = "b.md"
+    r2.chunk.id = "c2"
+    r2.score = 0.7
+
+    mock_ret = _make_mock_retriever(results=[r1, r2])
+    wrapper = AgentBenchRetriever(retriever=mock_ret, top_k=5)
+    docs = await wrapper.ainvoke("test")
+
+    assert len(docs) == 2
+    assert docs[0].page_content == "First"
+    assert docs[1].page_content == "Second"