Introducing DeepEval, RAGAS, and Phoenix Judges in MLflow
Summary
MLflow has integrated over 50 LLM-as-a-judge metrics from DeepEval, RAGAS, and Arize Phoenix directly into the MLflow Scorer API. Practitioners can now run and compare these third-party evaluations alongside native MLflow judges in a single API call and unified visual UI to debug and evaluate RAG pipelines and agents.
Summary from the captured archive. Follow the original source for full context.
More from MLflow Blog
5 Simple Tips to Supercharge Your Claude SDK Experience with MLflow
Claude SDKs now emit OpenTelemetry traces natively, allowing you to seamlessly monitor and evaluate your agent runs with MLflow. Learn five practical tips to capture unified end-to-end traces, track token usage, score executions, and compare performance across different prompts.
How to trace Claude agents via Native OpenTelemetry Export
Claude now natively emits OpenTelemetry traces across Claude Managed Agents, the Messages API, the Claude SDKs, and Claude Code. Learn how to configure the export for each product using open-source MLflow as your trace backend.
Can Jev replace your LLM judge? Part 2: Testing harder answers
TypeSafe's Jev model evaluates MLflow QA answers roughly seven times faster and significantly cheaper than larger LLMs, but it misses subtly incorrect technical details such as reversed roles and omitted permissions. Ahead of Jev integration arriving in MLflow 3.17, practitioners can use its confidence probabilities to fast-track clear verdicts while escalating borderline cases to a larger model for review.
Can Jev replace your LLM judge? Evaluating quality, cost, and latency
Benchmarking Jev against GPT, Claude, and DeepSeek using MLflow reveals how it measures up as an alternative LLM judge. Review side-by-side performance across quality, cost, and latency to see if Jev can replace your current evaluation models.
Evaluating and Improving Agent Skills with MLflow
MLflow enables evaluation-driven development for AI agent skills by combining traces, evaluation datasets, custom evaluators, and experiment tracking to treat agent capabilities like testable software components. Instead of grading only final outputs, teams can use execution traces to systematically measure critical behaviors such as tool sequencing, policy compliance, and execution efficiency across skill iterations.