Usage & Prompting
Grounding
Answers are tied to verifiable foundations: supplied documents, databases, search results, ideally with evidence of which statement rests on what. Grounding turns “the model says” into “the source says, the model summarizes.” For enterprise use it is less a feature than a prerequisite.
An example
Two instructions, one difference. “What does our travel expense policy say about the rail discount card?” lets the model answer from memory: it doesn't know your policy and still formulates something plausible. “Here is the text of the policy. Answer exclusively from it and quote the paragraph. If it isn't in there, say so” ties the answer to a document you can check.
A closer look
Answers are bound to supplied documents and given citations. The effect is twofold: the error rate drops, and errors become checkable in the first place, because the claim has an address.
Why it matters
Grounding is the most effective single measure against invented content and at the same time the one most often implemented halfway. Supplying the source is not enough: it takes the instruction to restrict itself to it, the obligation to give a passage, and a permitted non-answer. If the last part is missing, the model prefers to fill the gap rather than name it. The side benefit is organizational: when every answer carries a citation, an unverifiable statement becomes a transaction a human can approve in seconds.
Commonly misunderstood
A citation is not evidence. Models also cite passages that don't support the statement. The link itself has to be spot-checked.