databricks-mlflow-evaluation

MLflow evaluates GenAI applications through `mlflow.genai.evaluate()`, allowing you to run benchmark assessments using built-in scorers like Safety and Correctness, or apply custom scorers to measure groundedness and relevance.

Trace analysis via `mlflow.search_traces()` is the best way to debug agent regressions, enabling you to inspect detailed execution traces to identify exactly why your agent fails on specific queries.

Prompt optimization with MLflow GEPA works by using `mlflow.genai.optimize_prompts()` to automatically improve your prompts, analyzing evaluation results to iteratively tune and enhance LLM performance.

Yes, you can align LLM judges with domain expert preferences in MLflow using MemAlign, which calibrates automated evaluation judges to reflect specific domain expertise and human grading standards.

You set up production monitoring for live AI traffic by configuring continuous quality checks within the MLflow ecosystem, ensuring your deployed generative AI models maintain accuracy and safety over time.

Yes, MLflow GenAI evaluation supports custom scorers for RAG agents, allowing you to define specific metrics for groundedness and relevance alongside built-in safety checks to evaluate your retrieval augmented generation applications.