A QA team is preparing to release a Foundry-hosted assistant that answers questions using a retrieved knowledge base of internal manuals. Before release, they want an automated metric that scores whether each generated answer is actually supported by the documents it retrieved, so they can flag responses where the assistant states something the source material never said. Which evaluation metric fits this need?

The groundedness metric is the evaluation metric designed to check whether a generated answer is actually backed up by the retrieved source documents, which directly catches cases where the assistant states something the manuals never said, the exact failure mode this QA team wants to detect before release. The latency metric that measures how many milliseconds each response takes to generate is incorrect because it measures response speed, which says nothing about whether the content of the answer is factually supported by retrieved material. The cost metric that tracks the number of tokens consumed per response is incorrect because it tracks how much a response costs to generate, which is a budgeting concern unrelated to whether the stated facts trace back to the source documents. The content filter severity metric that logs how often unsafe categories are triggered is incorrect because it logs how often categories like violence or hate speech are triggered, which addresses harmful content rather than unsupported or fabricated factual claims. Since the team's concern is specifically that the assistant might assert things beyond what its retrieved documents actually contain, the metric built to compare generated claims against source text, groundedness, is the one that answers their question.