LangChain adapter

Auto-trace LangChain retrieval chains with no per-call code — one install() at startup.

import veralith.adapters.langchain as adapter
adapter.install()
 
# every RetrievalQA / RetrievalQAWithSourcesChain .invoke() now auto-logs

What it patches

The adapter wraps two chains at the class level, so all instances trace:

• RetrievalQA.invoke
• RetrievalQAWithSourcesChain.invoke

install() is idempotent; uninstall() restores the originals (handy in tests).

What it extracts

From each .invoke() call it pulls a full trace:

• Query — from the input dict (query / question / input) or a bare string.
• Response — from the result (result / answer / output_text / output).
• Context — from source_documents (falling back to context / sources), reading each document's page_content and its metadata.source / metadata.score.

If a LangChain version returns a shape it can't parse, it warns and passes the result through untouched — your chain never breaks.

The LCEL blind spot

Warning: the adapter only covers RetrievalQA and RetrievalQAWithSourcesChain. Chains built with LCEL (the prompt | llm | ... pipe syntax) or custom Runnables, as well as LlamaIndex and agent frameworks, are not auto-traced.

For those, call veralith.log() at the point where you have the query, the retrieved documents, and the final answer — see Recipes by code shape.