RAG

RAG joins two stages: retrieval or selection of passages from a defined corpus, followed by answer generation conditioned on those passages. The corpus may contain company documents, a knowledge base, or public web pages; its boundary comes from the implementation, not from the term itself.

Section sources:[1] Microsoft Learn[2] Anthropic Support

Conventional search can return documents without generation. Grounding describes an answer's support in evidence, while RAG describes a way to deliver external context to a model. Product web search may be part of RAG, but links alone are not enough to establish that.

Section sources:[1] Microsoft Learn[2] Anthropic Support

In an enterprise assistant, an employee query may retrieve five passages from an internal knowledge base and provide them to the model. If the answer exposes two links, that is a separate interface decision: retrieval may use five passages while attribution shows only two.

Verify the architecture through product documentation, application traces, or a controlled experiment with a known corpus. Record the query, index version, retrieved passages, generation prompt, and answer; without those artifacts, label RAG as a hypothesis.

RAG reduces dependence on the model's parametric knowledge but does not eliminate irrelevant retrieval, stale corpora, conflicting documents, or generation errors. It also does not guarantee that users will see links to the context used.