databricks-mlflow-evaluation

You can evaluate GenAI agents for safety and groundedness by running mlflow.genai.evaluate() to generate traces and review baseline scores across structured metrics within an end-to-end MLflow evaluation pipeline.

You can create evaluation datasets from production traces by leveraging MLflow utilities alongside Unity Catalog integration, enabling production monitoring and subject matter expert feedback loops for your GenAI agents.

The best way to align judges for GenAI evaluation workflows is using MemAlign, which provides structured judge alignment to ensure scorer accuracy and relevance within trace-aware evaluation pipelines.

Yes, you can optimize prompts automatically during MLflow GenAI evaluation by applying GEPA-based prompt optimization techniques to refine and enhance your GenAI agent performance iteratively.

Yes, MLflow evaluation supports retrieval-augmented workflows by providing trace-based analysis and scorer integration tailored to measure correctness and groundedness in retrieval-augmented GenAI agent pipelines.