Grounding
Forcing the answer to come from sources you supplied, with citations pointing at them, not from what the model vaguely recalls.
See it
What it is
Grounding is the contract that an answer must come from evidence you supplied, and must point at which piece it came from. Retrieval is the plumbing that puts documents in the prompt; grounding is the rule about what the model is then allowed to say. The two get conflated, but you can retrieve perfectly and still get an ungrounded answer that quietly blends the docs with whatever the model remembers from training.
In practice it is three moves. Number every chunk in the prompt and require inline citations like [3] on each claim. Ask for the supporting quote verbatim, not a paraphrase. Give an explicit escape hatch, because 'the sources do not cover this' is a correct answer and a model with no permission to say it will invent something instead. Then verify server side: check the cited id exists and the quoted span actually appears in that chunk.
Gotcha: a citation is not proof. Models produce confident, well-formatted references to the wrong chunk, or cite a real chunk that does not support the sentence attached to it. Unverified citations are worse than none, because they buy trust the answer has not earned. Grounding lowers the hallucination rate; it does not zero it.
Ask AI for it
Rewrite this answering prompt to be strictly grounded. Inject the retrieved chunks as a numbered SOURCES block, each with an id, title and URL. Instruct the model to answer using only those sources, to append a bracketed source id to every factual sentence, and to reply exactly 'Not covered by the provided sources' when the answer is not there rather than filling the gap from memory. Have it return JSON with answer, citations (array of {source_id, quoted_span}), and unsupported_claims. Then add a validation step that drops the answer when a cited id is missing or the quoted span does not appear verbatim in that source.