Retrieval-augmented generation

Retrieval-augmented generation, or RAG, is the architecture behind most of the "the model looked it up" claims you will read this year. It pairs a language model with a search step over a defined document corpus. The model writes its answer using the retrieved passages as context.

RAG reduces, but does not eliminate, the risk of fabrication. The model can still ignore the retrieved passages, summarise them incorrectly, or fill gaps with confident invention. The retrieval step itself can also miss the relevant document and surface a near-match that supports a wrong conclusion.

For newsroom use, the question is not whether a system uses RAG. It is which corpus is being searched, who curates it, and whether the answer cites the specific passages it relies on in a form an editor can verify.