Groundedness

Groundedness is the degree to which the claims in a generated answer are supported by the source material the model was given. In retrieval-augmented generation, it separates two questions that a wrong answer mixes together: did the system find the right evidence, and did the answer stay within what that evidence says?

Google’s grounding-check documentation requires complete support for each claim and treats partial support as insufficient. Its example is a sentence with the right founders and the wrong founding year: half right is not grounded. The service returns an overall support score between 0 and 1 and cites the source passages behind each claim, and an application can set a threshold below which answers are withheld.

Checking it

Assess support for individual claims and aggregate it under a stated answer-level rubric. Take a leave-policy passage stating that employees may carry over five days into the next year, and the answer You can carry over five days, and unused days are paid out in March. The first claim is supported only if the questioner’s employee status and the rule’s applicability are established; the payout claim is not supported by that passage. A useful answer format asks the model to cite a passage identifier for each claim, and the application then checks that each cited passage exists, was actually supplied, and semantically supports the full claim, including its subject, scope, time, and conditions. Literal presence alone is insufficient. A citation to a real passage that does not support its claim is worse than no citation, because it lends the claim false authority.

Automated checks, whether a dedicated service or a second model acting as judge, are themselves models and make mistakes; their scores should be validated against human judgments on a sample before they are used to block answers. When support is missing, the honest responses are to answer only the supported part, to state that the available authorized evidence is insufficient, or to ask for clarification, rather than to fill the gap. Further retrieval may find missing evidence. Failure to find support in selected passages does not establish absence from the whole corpus; define that larger scope before labeling a question unanswerable, and avoid revealing inaccessible sources through the response.

Grounded is not the same as correct

An answer can faithfully summarize an outdated passage while failing to answer a current-policy question correctly. But asserting that an old rule applies today adds a temporal claim that may lack support. Likewise, a statement about employees does not establish a contractor’s entitlement. Evidence quality and claim support are distinct, but changing audience, conditions, or time can make the claim itself ungrounded. Groundedness measures the step from evidence to answer; whether the evidence was the right evidence is a question about retrieval, freshness, and permissions. An unsupported answer, by contrast, is a hallucination in the usual sense, even when it happens to be true. How the two kinds of failure are told apart is covered in Last Year’s Policy, Cited Correctly.

References

• Google Cloud, Check grounding with RAG (Creative Commons Attribution 4.0).