RAG

RAG, retrieval-augmented generation, retrieves evidence first, then generates an answer from that evidence instead of parametric memory alone.

A RAG system has a retriever (vector search, BM25, or hybrid) and a generator (usually an LLM). Production stacks add chunking, metadata filters, reranking, citations, tenant access control, and faithfulness evals. RAG reduces hallucination only when retrieved evidence is relevant and the model is required to use it. Empty retrieval plus a chatty model is still an ungrounded answer.