Evaluation
Grounding
Grounding is the property of an answer being supported by supplied source material rather than by the model’s training data.
A grounded answer can be checked: every claim traces to a passage that was in the prompt. That checkability is what makes retrieval-augmented systems auditable in a way a bare model call is not.
Checking it mechanically is imperfect — lexical overlap scores a correct paraphrase low and a fluent fabrication high — but the numeric check is reliable: a figure absent from the context was invented.
In practice
A citation is not evidence of grounding. Models produce citation-shaped text for sources they did not use, so the check that matters is verifying the cited chunk is actually in the retrieved set — cheap to implement and the difference between a citation and a decoration. Faithfulness failures survive every other check because they look correct.
Common questions
How do I check whether an answer is grounded?
Spot-check claims against the retrieved chunks. Faithfulness failures survive every other check because they look correct — well-formed, well-cited, and not actually supported by the source they point at.
Should the model use its own knowledge alongside retrieval?
Decide explicitly and say so in the prompt. Left unspecified it will mix the two, and afterwards you cannot tell which sentence came from your documents and which came from training.