RAG (Retrieval-Augmented Generation) is the dominant data-grounded answering pattern: before answering, the most relevant chunks from your knowledge base are retrieved and injected into the prompt. The model works on your live, real text instead of frozen memory.
The standard loop: chunk documents → embed → store in a vector DB → retrieve on the user query → (rerank) → inject → generate with citations. The 2026 generation reaches GraphRAG and Agentic RAG: multi-hop, self-questioning retrieval.
When to skip: small static docs (paste them) and tasks needing no external data. RAG carries infrastructure cost; its value scales with data volume and freshness.
