Retrieval-Augmented Generation
Also called: RAG, grounded generation
Retrieving relevant source documents first, then generating an answer constrained to those sources, so output can be traced back to a citation.
Retrieval-augmented generation splits the work in two. A retrieval step finds documents relevant to the question, then the generation step answers using only those documents. The answer can then carry citations to the retrieved sources.
This is the architecture behind clinical reference tools that answer with references attached, and it is the main practical mitigation for hallucination in clinical question answering.
Where This Gets Misread
Grounding reduces fabrication, it does not eliminate it, and it introduces a second failure point. If retrieval misses the relevant paper, the model answers confidently from what it did retrieve, and the resulting answer looks exactly as well-cited as a complete one. A citation proves a source exists, not that it supports the claim attached to it.