Grounding
Grounding means tying a model's answer to specific, supplied sources, such as retrieved documents or database results, so that each claim can be traced back and checked.
Also known as: grounded generation
An ungrounded answer comes from whatever the model absorbed in training, with no way to see where a claim came from. A grounded answer is built from material put in front of the model at the time of the question, usually by retrieval, and ideally cites which source supports which claim. That gives a reader something to check and gives the model less reason to invent.
Grounding reduces hallucinations but doesn’t guarantee accuracy. The sources can be wrong or stale, the retrieval can miss the right document, and the model can still go beyond what the sources say. Teams measure the last of these with faithfulness checks, and many enterprise hallucination problems turn out to be data problems.
