Retrieval-Augmented Generation (RAG)

RAG retrieves relevant information from an external knowledge source and supplies that information as context to a generative model. It is useful when answers need information that is private, frequently changing, or outside the model's training knowledge.

Core flow

Documents → parse/chunk → embeddings/index → query → retrieve/rank → context → model → answer

Why use RAG?

• Ground responses in a controlled source of information.
• Update knowledge without retraining the foundation model.
• Return citations or source references when the retrieval system preserves provenance.
• Apply access controls to retrieved information.

RAG is not automatically accurate

A RAG application can still produce unsupported answers if retrieval is poor, context is incomplete, authorization is wrong, or the model ignores the evidence. Evaluate retrieval and generation separately.

Production checklist

• Version ingestion and chunking rules.
• Use an embedding model appropriate for the data and language.
• Measure retrieval quality on representative queries.
• Apply document-level authorization where required.
• Monitor stale indexes and ingestion failures.
• Evaluate groundedness and answer quality.

Connected knowledge

embeddings → vector-search troubleshooting → RAG hallucinations → production RAG architecture.