Retrieval-Augmented Generation (RAG)

Retrieval-Augmented Generation (RAG) retrieves external material at request time and supplies selected results to a generative model.

A RAG system indexes or queries a source collection, retrieves candidates for a request, and includes selected passages or records in the model context. Retrieval may use keywords, vectors, database filters, graphs, APIs, rerankers, or combinations; embeddings and fixed-size chunks are common but not required. The model can use information that was not present in its training data without changing its weights.

Builder example

RAG can connect answers with private or changing sources when ingestion and retrieval are current. Failures can begin in parsing, metadata, indexing delay, query construction, ranking, source authority, or generation. Evaluate retrieval recall and answer support separately, preserve source dates and identifiers, and show what evidence the answer used.

An employee asks about parental leave. The system retrieves last year's handbook section because it is semantically similar to the current one, and the model gives a confident wrong answer.

Filter by document freshness, prefer the current policy as the source of record, and cite the specific section so the user can verify.

Common confusion: Retrieval does not guarantee that the answer uses the right passage or remains faithful to it. A system can retrieve correct evidence and still contradict it, or retrieve a plausible but stale near-match.