Retrieval-Augmented Generation (RAG)

Retrieval-augmented generation (RAG) is a technique in which an AI system first retrieves the most relevant passages from approved knowledge sources, adds them to the prompt as context, and then generates an answer grounded in that content. It lets an agent answer accurately about an organization's policies and data without retraining the model.

Grounding is the goal; RAG is the mechanism. Retrieval works on meaning rather than keywords: the question is converted into an embedding and matched against a vector index, which is why "time off" finds the leave policy. Answer quality is set by source quality, so clean, well-chunked, well-labeled documents retrieve far better than one large undifferentiated file. The misconception is that RAG teaches the model your data permanently. It does not; it fetches the relevant content fresh for each question.