---

Why langchain-longprobe?

Every RAG developer faces the same problem: you upgrade LangChain, swap a vector store, or tweak a chunking strategy — and your retrieval silently degrades. Existing evaluation tools (Ragas, DeepEval) tell you the LLM answer got worse, but they can't tell you which specific chunks were lost.

langchain-longprobe bridges this gap:

Feature	LangChain Evaluators / Ragas	langchain-longprobe
Focus	LLM response quality	Retrieval stability
Speed	10s–60s (LLM-as-judge)	Sub-second (chunk match)
Feedback	Pass/Fail score	Visual diff of lost/gained chunks
Workflow	Batch analysis	pytest / CI integration
Detects	Bad answers	Why retrieval failed

Installation

pip install langchain-longprobe

This installs both langchain-longprobe and the core longprobe library.

Quick Start

1. Define Golden Questions

Create a goldens.yaml with your expected retrieval results:

name: "my-rag-golden-set"
version: "1.0"

questions:
  - id: "q1"
    question: "What is the refund policy?"
    match_mode: "id"
    required_chunks:
      - "chunk_refund_01"
      - "chunk_refund_02"
    top_k: 5

  - id: "q2"
    question: "What are the payment terms?"
    match_mode: "text"
    required_chunks:
      - "net 30 days from invoice"
    top_k: 5

2. Probe Your Retriever

from langchain_longprobe import RetrievalProbe
from langchain_community.vectorstores import Chroma

# Your existing LangChain setup
vectorstore = Chroma(persist_directory="./db")
retriever = vectorstore.as_retriever(search_kwargs={"k": 5})

# Create a probe
probe = RetrievalProbe.from_retriever(
    retriever=retriever,
    goldens_path="goldens.yaml",
)

# Run regression check (sub-second)
report = probe.run()
print(f"Recall: {report.overall_recall:.2%}")
print(f"Pass Rate: {report.pass_rate:.2%}")

# See exactly what's missing
for qid, chunks in probe.get_missing_chunks().items():
    print(f"  {qid}: lost {chunks}")

3. Track Regressions Over Time

# Save a baseline after your first successful run
probe.run()
probe.save_baseline("v1.0")

# After code changes, compare against baseline
probe.run()
diff = probe.diff("v1.0")
print(f"Regressions: {len(diff['regressions'])}")
print(f"Improvements: {len(diff['improvements'])}")

API Reference

RetrievalProbe — Main Entry Point

The recommended way to use langchain-longprobe. Wraps any LangChain BaseRetriever.

from langchain_longprobe import RetrievalProbe

probe = RetrievalProbe.from_retriever(
    retriever=your_retriever,
    goldens_path="goldens.yaml",
    recall_threshold=0.85,
)

report = probe.run()
probe.save_baseline("v1.0")
diff = probe.diff("v1.0")
missing = probe.get_missing_chunks()

ProbedRetriever — Drop-in Retriever Wrapper

A LangChain BaseRetriever that wraps your retriever and adds regression testing. Use as a drop-in replacement in your existing chains.

from langchain_longprobe import ProbedRetriever

probed = ProbedRetriever(
    retriever=your_retriever,
    goldens_path="goldens.yaml",
    check_on_invoke=False,  # set True for automatic checks
)

# Use exactly like a normal retriever
docs = probed.invoke("What is the refund policy?")

# Manually trigger a regression check
report = probed.check()
probed.save_baseline("v1.0")

LongProbeCallbackHandler — Passive Monitoring

Attach to any LangChain chain to passively record retrieval events.

from langchain_longprobe import LongProbeCallbackHandler

handler = LongProbeCallbackHandler(
    goldens_path="goldens.yaml",
    recall_threshold=0.85,
    fail_on_regression=True,
)

# Attach to retriever calls
docs = retriever.invoke("query", config={"callbacks": [handler]})

# Inspect results
print(handler.retrieval_log)

# Run a full check
report = handler.run_probe(retriever)

RetrievalRegressionRunnable — LCEL Integration

Use LongProbe as a composable step in LangChain Expression Language chains.

from langchain_longprobe import RetrievalRegressionRunnable

runnable = RetrievalRegressionRunnable(
    retriever=your_retriever,
    goldens_path="goldens.yaml",
    fail_on_regression=True,
)

# Invoke with options
result = runnable.invoke({
    "top_k": 10,
    "save_baseline": "v2.0",
    "baseline_label": "v1.0",  # compare against this
})

print(result["overall_recall"])
print(result["missing_chunks"])

Pytest Integration

conftest.py

import pytest
from langchain_longprobe import RetrievalProbe

@pytest.fixture
def probe(my_retriever):
    return RetrievalProbe.from_retriever(
        retriever=my_retriever,
        goldens_path="goldens.yaml",
        recall_threshold=0.85,
    )

Writing Tests

def test_retrieval_recall(probe):
    """Ensure retrieval recall stays above threshold."""
    report = probe.run()
    assert report.overall_recall >= 0.85, (
        f"Recall dropped to {report.overall_recall:.2f}. "
        f"Missing: {probe.get_missing_chunks()}"
    )

def test_no_regression(probe):
    """Ensure no chunks were lost vs. baseline."""
    probe.assert_no_regression("v1.0")

Command Line

pytest --langchain-longprobe-goldens goldens.yaml --langchain-longprobe-threshold 0.85

GitHub Actions

name: RAG Regression Check

on: [push, pull_request]

jobs:
  rag-probe:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: actions/setup-python@v5
        with:
          python-version: "3.12"
      - run: pip install langchain-longprobe
      - name: Run regression check
        run: pytest tests/test_rag_regression.py -v
        env:
          OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}

Examples

Basic Regression Check

from langchain_longprobe import RetrievalProbe
from langchain_openai import OpenAIEmbeddings
from langchain_community.vectorstores import Chroma

embeddings = OpenAIEmbeddings()
vectorstore = Chroma(
    persist_directory="./chroma_db",
    embedding_function=embeddings,
)
retriever = vectorstore.as_retriever()

probe = RetrievalProbe.from_retriever(retriever, goldens_path="goldens.yaml")
report = probe.run()

if report.regression_detected:
    print("⚠️  Regression detected!")
    for qid, chunks in probe.get_missing_chunks().items():
        print(f"  Question {qid}: lost chunks {chunks}")
else:
    print("✅ All chunks present")

CI/CD Pipeline Integration

# tests/test_rag_regression.py
import pytest
from langchain_longprobe import RetrievalProbe

@pytest.fixture(scope="session")
def probe():
    from langchain_community.vectorstores import Chroma
    retriever = Chroma(persist_directory="./db").as_retriever()
    return RetrievalProbe.from_retriever(
        retriever=retriever,
        goldens_path="goldens.yaml",
        recall_threshold=0.85,
    )

def test_recall_above_threshold(probe):
    report = probe.run()
    assert report.overall_recall >= 0.85

def test_no_regressions_vs_baseline(probe):
    probe.assert_no_regression("production")

def test_critical_questions_pass(probe):
    report = probe.run()
    for result in report.results:
        if "critical" in result.question_id:
            assert result.passed, f"Critical question {result.question_id} failed"

Part of the Long Suite

langchain-longprobe is the official LangChain integration for LongProbe, part of the EnDevSols Long Suite of RAG tools:

• LongParser — Document ingestion and chunking
• LongTrainer — RAG chatbot framework
• LongTracer — Hallucination detection
• LongProbe — Retrieval regression testing (core library)
• langchain-longprobe — LangChain integration ← You are here

Contributing

We welcome contributions! Please see the LongProbe Contributing Guide for guidelines.

License

MIT License — see LICENSE for details.

---