AI-103 Practice Question 16: Groundedness Evaluator Choice
Microsoft Azure AI Apps and Agents Developer Associate (AI-103) · Test 2 · question 6 of 10
Before launch, reviewers of a RAG assistant flag a recurring problem: the answers read fluently, address the question that was asked, and cite the right policy area, but they contain premium figures and effective dates that appear nowhere in the retrieved chunks. Spot-checking confirms the retrieved chunks themselves are the correct ones. The test set holds a query, the concatenated retrieved context, and the produced response for each case; nobody has written ideal reference answers.
Which built-in evaluator should gate the release?
Show answer and explanation
Answer:
Groundedness is the system evaluator for fabrication: Microsoft describes it as the *precision* aspect of the response; the answer contains nothing outside the grounding context. It needs response and context (with query optional but recommended), which is exactly the data this team already has, and it returns a 1–5 score with a default pass threshold of 3.
- A is wrong because Relevance answers "does the response address the query?", and the reviewers confirm it does. A fluent answer built on invented numbers can score well on relevance, which is why relevance alone never catches this failure. - B is wrong because Retrieval is a *process* evaluator on the upstream step, scoring how relevant the retrieved chunks are to the query. The team has already established that retrieval is returning the right chunks. - D is wrong on two counts: Response Completeness measures *recall* (whether the answer misses critical information) which is the opposite of the observed problem, and it requires ground_truth, which this team does not have.
Reference: Retrieval-Augmented Generation (RAG) evaluators for generative AI