Retrieval-Augmented Generation

A typical RAG pipeline prepares documents, indexes searchable representations, retrieves candidates for a query and passes selected evidence to a generator. Retrieval may be lexical, vector-based, graph-assisted or combined. The original RAG research integrated retrieved passages with a generative model; deployed applications often use a simpler explicit retrieve-then-prompt architecture. RAG differs from fine-tuning because adding or changing evidence in the collection does not inherently update the model's parameters. It also differs from merely providing a long document: retrieval chooses relevant evidence from a larger source space for each request.

The practitioner builds ingestion with source versioning, chooses retrieval and context policies and evaluates questions with known supporting evidence. Tests distinguish failure to retrieve a passage from failure to use it correctly. Answer checks inspect coverage, unsupported claims and citations. Useful artifacts include the indexing pipeline, relevance labels, generation contract and traces of selected passages. Operational work covers access controls, document updates and deletion. The application needs an abstention or clarification path when the collection cannot support a reliable answer.

RAG can reduce some knowledge gaps but does not eliminate hallucination. Retrieval may miss the evidence, return an unauthorized document or rank a near match above the correct item. The generator may ignore qualifiers or combine conflicting passages incorrectly. Quality depends on the collection and complete pipeline, so adding a vector database alone does not establish a trustworthy system. Grounding, source reliability and access enforcement require their own checks.