AI & Agent Evaluation
2,151total visitsadmin

MLflow optimize_prompts() / GepaPromptOptimizer

product docs · source date 2026-05-01 · added 2026-08-07 15:52:43 · updated 2026-08-07 15:52:43 · Open original blog

Problems / challenges / motivations

  • Prompt optimizers mostly live inside one framework, so adopting them means adopting the framework.
  • Optimized prompts are artifacts that need versioning, diffing, and rollback like any other deployed asset.
  • Without CI integration, optimization is a one-off experiment rather than a maintained process.

Key ideas

  • `mlflow.genai.optimize_prompts()` ships GEPA (via `GepaPromptOptimizer`) as a first-class, framework-agnostic prompt optimizer alongside Metaprompting.
  • Supports custom scorers and judges, joint multi-prompt optimization, and model switching.
  • Automatically registers optimized prompts as versioned artifacts in the Prompt Registry, with diffs, tracking, and rollback.

Why it matters for AI engineering

  • This brings GEPA into a mainstream MLOps platform and works across LangChain/LangGraph, CrewAI, AutoGen, the OpenAI Agents SDK, and Pydantic AI.
  • It is the productionization layer that turns reflective optimization into a governed, CI-friendly workflow — a bigger distribution signal than the method's adoption inside DSPy alone.
  • Joint multi-prompt optimization matters because in a compound system, optimizing one module's prompt in isolation is how you get local wins and global regressions.
  • Caveat: exact release date unverified. Vendor product docs, and the underlying GEPA gains carry the same author/vendor-reported caveats, including VISTA's demonstration that GEPA can degrade a defective seed prompt.

Comments

No comments yet.