Implement Self-RAG for Enhanced Document Analysis

An archived LangGraph example (no longer maintained - see LangChain's consolidated documentation for current guidance) implementing Self-RAG: a RAG strategy built around four self-reflection decisions from the Self-RAG paper. It decides (1) whether retrieval is even needed for the question, (2) whether each retrieved chunk is relevant to the question, (3) whether the generation from each chunk is fully, partially, or not supported by that chunk (checking for hallucination), and (4) how useful the generation is as a response to the question, scored 1-5.

Use this as a reference implementation for a RAG pipeline that self-grades at every stage - retrieval necessity, document relevance, groundedness of the generation, and answer usefulness - rather than a single-pass retrieve-and-generate flow. It is explicitly archived and no longer updated, so treat it as a study pattern rather than a production dependency - check LangChain's current consolidated documentation for up-to-date APIs.

Decision 1 takes a question (and optionally a prior generation) and outputs yes/no/continue on whether to retrieve. Decision 2 takes the question plus each retrieved chunk and outputs relevant/irrelevant. Decision 3 takes the question, chunk, and generation, and outputs fully supported/partially supported/no support. Decision 4 takes the question and generation and scores usefulness 1-5. The example indexes three blog posts as its retrieval source and builds the full decision flow as a LangGraph graph.

Built with LangGraph and LangChain, with LangSmith recommended for tracing and debugging - the source links two public LangSmith trace examples showing the graph's actual run-time behavior.

Developers building a RAG pipeline that needs explicit self-grading at each stage - whether to retrieve, whether retrieved content is relevant, whether the generation is grounded, and whether the answer is actually useful - instead of trusting a single retrieve-and-generate pass.