Grounding: connect each claim to something checkable

Grounding connects a model's output to information outside its unsupported generation: retrieved documents, database results, measurements or other verifiable sources.

Retrieval is one way to supply evidence. The harder question is whether the generated claim actually follows from the evidence supplied.

A citation is only the start

Suppose a retrieved passage says a service is available in three countries. A response that cites it while claiming worldwide availability has a real source but an unsupported conclusion.

Separate three checks: did the system find relevant evidence, did it represent that evidence accurately, and does it support the particular claim? A link can establish provenance without establishing support.

In a useful grounded workflow, source IDs survive retrieval and generation so a reader or checker can trace a statement back to the relevant passage. Conflicting or missing evidence should change the answer rather than disappear behind fluent wording.

Grounding improves access to external information; it does not turn source material into trustworthy instructions or make an incorrect source true.