LangChain Integration¶
LongTracer hooks into LangChain's callback system to automatically capture retrieval, LLM, and verification spans — no changes to your chain required.
Install¶
Usage¶
from langchain.chains import RetrievalQA
from langchain_community.vectorstores import Chroma
from longtracer import LongTracer, instrument_langchain
# 1. Init LongTracer
LongTracer.init(verbose=True)
# 2. Build your chain as normal
chain = RetrievalQA.from_chain_type(
llm=your_llm,
retriever=your_vectorstore.as_retriever()
)
# 3. Instrument — one line
instrument_langchain(chain)
# 4. Use your chain as normal — verification happens automatically
result = chain.invoke({"query": "What is the capital of France?"})
What Gets Captured¶
| Span | What it records |
|---|---|
retrieval |
Retrieved chunks, count, latency |
llm_prep |
Prompt text, context length |
llm_call |
LLM answer, model name, latency |
eval_claims |
Per-claim verification results |
grounding |
Trust score, hallucination count, verdict |
Viewing Results¶
Notes¶
- Works with any LangChain chain that uses a retriever (
RetrievalQA,ConversationalRetrievalChain, custom chains) - Verification is triggered at the end of the root chain, after the LLM has responded
- If no chunks are retrieved, verification is skipped gracefully
- A failing verification never crashes your chain — all errors are logged as warnings