databricks-mlflow-evaluation

MLflow GenAI evaluation systematically quantifies agent performance using domain-specific scorers and ground-truth datasets. You prepare an evaluation dataset and run mlflow.genai.evaluate() to start assessing outputs against aligned standards.

Yes, MLflow GenAI evaluation supports Unity Catalog trace ingestion. This allows you to ingest production traces directly, enabling continuous scoring and monitoring of deployed agents against your aligned evaluation standards.

The best way to compare prompt variants is using a common evaluation dataset to run systematic MLflow GenAI evaluations. This tracks improvements by applying domain-specific scorers and ground-truth datasets to each variant.

Yes, MLflow GenAI evaluation integrates GEPA-based prompt optimization. This automates the refinement of agent prompts by aligning domain-specific scorers and MemAlign-based judge alignment within end-to-end evaluation pipelines.

You align LLM judges using MemAlign-based judge alignment workflows. This ensures the automated scorers in your MLflow GenAI evaluation pipelines accurately reflect domain-specific criteria and ground-truth expectations.

You need a prepared ground-truth evaluation dataset containing domain-specific scenarios. The skill supports dataset creation workflows, allowing you to orchestrate scorers and apply automated evaluation pipelines across Databricks environments.