Rag

Retrieval-augmented generation grounds model answers with chunks from vector or keyword indexes before decoding tokens. Chunking, hybrid search, and rerankers control relevance while citations reduce hallucinations. Refresh embeddings whenever documentation meaningfully changes.