Exam Databricks-Generative-AI-Engineer-Associate Topic 4 Question 91 Discussion

In the Agent Evaluation terminology used by this question, correctness compares the generated response with the expected answer, while context_sufficiency determines whether the retrieved context contains enough information to produce that expected answer. Both therefore depend on ground-truth information supplied through expected_response. Groundedness instead examines whether the response is supported by the retrieved context; it does not inherently need a reference answer. Chunk relevance and relevance to the query assess alignment with the user's request, while guideline adherence evaluates compliance with specified instructions. Databricks explicitly distinguishes the ground-truth-dependent judges in its migration documentation. In MLflow 3, names and schemas have evolved, including expectations.expected_response and explicitly selected scorers, so the question should be interpreted using its stated legacy judge terminology.
Databricks documentation