Grounding

Constraining and connecting a model result to external evidence, records, or observations relevant to the task.

A grounded claim identifies evidence that supports it, such as a policy passage, database record, calculation, sensor result, or tool output. 'The return window is 30 days' is grounded only if the cited current policy states that rule for the applicable case. Citation display, retrieval, and support checking are separate mechanisms; an answer can cite a real source that does not entail the claim.

Builder example

Grounding makes evidence paths inspectable and supports updates when a source changes. The workflow should record source identity, version, passage, and applicable conditions for decision-relevant claims. Creative or subjective tasks may need examples and constraints rather than factual citations.

You ask about your refund policy. The model gives a plausible-sounding answer based on what refund policies usually look like, not your actual policy.

Retrieve the actual policy document, cite the section, and make it easy for the user to verify by linking to the source.

Common confusion: Retrieving documents and grounding answers are two separate steps. A system can retrieve the right document and still generate an answer that ignores or contradicts it. Grounding requires verifying that the output actually reflects the retrieved evidence.