A publisher's support chatbot pulls passages from official documentation before answering. Evaluators notice that some responses stick precisely to what the retrieved passage says, while others quietly add plausible-sounding details that are not present in the passage at all, even though the retrieval step worked correctly both times. Which evaluation metric is specifically designed to score whether a generated response is supported by its retrieved source content?
Select an answer to reveal the explanation.
Short Explanation
There is a very specific kind of mistake happening here: the system found the right reference material, but sometimes it does not stick to it, sneaking in extra details that sound right but were never actually in that material. That is a distinct failure from writing badly or being disorganized, so the metric that catches it has to compare the response directly against the source passage and ask whether every claim in the response can actually be traced back to it. A score about how smoothly the sentences read would happily give full marks to a beautifully written fabrication. A score about logical structure cares about whether ideas connect sensibly, not whether they are true to a source. And a score built around comparing to some reference answer is answering a different question than 'does this match the passage it was supposedly built from.' The one built for exactly this check is the one that measures faithfulness to the retrieved content itself.
Full Explanation
The correct answer is C. Groundedness is the evaluation metric built specifically to measure whether a generated response's claims are actually supported by the retrieved or provided source content, which directly captures the failure being described: passages retrieved correctly, but the model sometimes adding claims not present in them. Option A is incorrect because fluency measures how natural and grammatically well-formed the language is, and a response can read perfectly smoothly while still fabricating unsupported details, so it would not catch this problem. Option B is incorrect because coherence measures whether a response is logically organized and internally consistent, which says nothing about whether its content matches the source passage it was supposed to be grounded in. Option D is incorrect because similarity typically compares a generated response against a reference answer for overlap in wording or meaning, not against the retrieved source passage for factual support, so it targets a different comparison than the one needed here.