An ungrounded model answers from patterns learned during training, which can be out of date or simply wrong for a specific company. A grounded system instead looks up current, approved information, such as a policy document or a product spec, before writing its answer, and it can show the source alongside the answer so a person can check it.
Retrieval-augmented generation is the most common way to ground a model, but grounding also includes simpler cases, such as passing today's account balance directly into a prompt instead of letting the model guess. Grounding reduces the rate of hallucination significantly, though it does not remove the risk completely, since the retrieved information itself can be outdated or incomplete.