Groundedness

Groundedness measures the relationship between a model's response and its source material, such as provided documents, retrieved passages, or conversation history, by asking whether every claim in the output can be traced back to that material. An ungrounded response might still be factually true, but if it is not supported by the given context, it is still a groundedness failure, since the model should have said it did not know or asked for more information instead of asserting a claim it could not back up.

This is closely related to but broader than fact-checking. Fact-checking verifies specific claims against real-world truth, while groundedness verifies claims against the specific context the model was given, which matters enormously for RAG systems where the whole point is to answer from retrieved documents rather than parametric memory.

Ungrounded responses are one of the most common root causes flagged in hallucination evaluation, and annotators checking groundedness are expected to compare the response against the source text directly rather than relying on their own general knowledge of the topic.