Before releasing a Foundry-hosted assistant that answers from a retrieved knowledge base, a QA team wants an automated, repeatable way to measure how often the assistant's answers are actually supported by the retrieved passages, without manually reading every response. What should they use?
Select an answer to reveal the explanation.
Short Explanation
What the team really wants is a score they can run again and again without a human reading every transcript, and that is exactly what the built-in evaluation tooling in Foundry is for: point it at a set of test questions and the passages that were retrieved for them, and it reports back how well each answer actually sticks to what those passages say. Checking a handful of conversations by hand after the product is already live catches problems too late and does not scale. Turning up the filters that catch hateful or violent language is a completely different safeguard aimed at harmful content, not at whether an answer's facts are backed up. And simply feeding the model more retrieved passages might change what it has to work with, but it will not tell anyone, by itself, whether the final answer is actually grounded in that material.
Full Explanation
The correct answer is D. Azure AI Foundry includes built-in evaluation metrics, including a groundedness metric, that can be run automatically against a test dataset of questions and retrieved passages to score how well each generated answer is supported by the source material, giving the QA team a repeatable, scalable measurement instead of manual review. Option A is incorrect because manual spot-checking after launch is neither automated nor repeatable, and it happens after users are already affected rather than during pre-release testing. Option B is incorrect because content safety severity thresholds control filtering of harmful categories like hate or violence; they say nothing about whether an answer's claims are actually supported by retrieved documents. Option C is incorrect because returning more retrieved passages might change what information is available to the model, but it does not by itself measure or report how grounded the resulting answers are; that still requires an evaluation step.