Retrieval-augmented generation

RAG is why a page still has to be reachable at the moment the question is asked. A fact that existed only in a training corpus can be stale, refused, or never attached to the company name. A fact on a fetchable URL can be retrieved for that question.

The generator writes from the passages it was handed. If those passages bury the answer, the model has to infer, and inference is where names get dropped. An extractable passage is a passage that does not ask the model to infer the point.

Not every answer product is RAG on every question. Some answers are memory, some are live retrieval, some are both. Publishing for retrieval covers the case you can actually influence.