Retrieval-augmented generation (RAG)

Also written RAG

The pattern where a model searches for relevant documents first, then writes its answer using what it found.

Without retrieval, a model answers from what it absorbed in training, which is fixed at a point in time and cannot be corrected. With retrieval, it fetches current documents and writes from those, which is why assistants can discuss things that happened after their training cut-off.

For anyone trying to be recommended, RAG is the mechanism that makes the work possible. If answers came only from training data, the only lever would be being famous enough to be memorised. Because retrieval happens at question time, a page published this month can be cited this month.

It also explains why citations skew toward third-party sources. The retrieval step is a search, and a search for the best tool in a category surfaces comparisons and reviews rather than vendor homepages.

GroundingTying a generated statement to a specific retrieved source, so the claim can be attributed rather than invented.Fan-out queriesThe hidden sub-questions a generative engine generates from one user prompt, each retrieving its own sources before the answer is written.Citation rateThe proportion of AI answers that cite at least one page on your own domain as a source.