Retrieval-augmented generation

A method that supplies a language model with retrieved documents so its answers are grounded in specific text rather than in whatever it absorbed during training.

For analysis work the benefit is traceability. A grounded answer can cite the paragraph it came from, which turns a plausible summary into something checkable.

It reduces invention without eliminating it. A model can still misread a retrieved passage, and the retrieval itself can return the wrong document.

Grounding is a control, not a guarantee. Verify against the source.