A Generative AI Engineer has deployed a customer-support agent in production that retrieves product documentation...

Option B preserves the relationships between each user query, generated response, and expert assessment. That context is necessary to turn production feedback into reusable evaluation cases. The engineer can curate representative examples from the Delta table, map them into MLflow’s evaluation schema, and retain the dataset for consistent comparisons between agent versions. Where correctness scoring is required, experts should provide corrected answers or expected facts; a star rating alone is not a ground-truth answer. Keeping only written comments loses essential request context. Selecting only five-star interactions biases the dataset and excludes failure cases that future versions should improve. Deploying highly rated responses is not a substitute for evaluating an updated agent against a stable, representative benchmark. Databricks documentation