Assess factual consistency of generated responses with trusted sources or knowledge bases.
Tests whether each statement in a generated output is actually supported by the source material the system was given. Distinct from asking whether a statement is true in the world: the question is whether the retrieved context entails it. Typically implemented with a natural-language-inference model scoring each claim against each retrieved passage.
Any retrieval-augmented system whose output a person will act on. Essential wherever the generated text becomes a record rather than a convenience.
Creative or exploratory generation, where there is no source to be grounded in and the check has nothing to score against. Also weak where the answer requires synthesis across passages, since entailment against any single passage will fail on a correct conclusion.
Groundedness scores must be surfaced to the reader, not merely logged. A system that computes the score and displays the answer regardless has bought monitoring, not safety.
Scoring the presence of a citation rather than its support. A statement citing a real passage that does not actually say what is claimed passes a naive check, and this is the most common failure mode in production.
A reconciliation narrative asserting that a variance is explained by three specific reconciling items. The checker verifies each item genuinely appears in the retrieved reconciliation record, so a fabricated fourth item is caught before a reviewer signs it.