dspy-gepa-reflective

GEPA optimizes agentic DSPy workflows by reflecting on full execution traces to search for Pareto-optimal improvements. Define a metric returning (score, feedback), configure a reflection LM, set the auto mode, and compile your agent to apply reflective optimization across multi-step processes with tool use.

Yes. GEPA is designed for complex agentic systems with tool use and textual feedback on failures. It processes full execution traces to identify improvement opportunities, making it suitable for agents requiring instruction-only optimization without labeled training data.

GEPA uses LLM reflection on complete execution traces combined with Pareto-based evolutionary search. Unlike instruction-only approaches, it leverages rich failure feedback and multi-step workflow context to optimize agent behavior across competing objectives simultaneously.

Define a metric callable that returns (score, feedback) pairs—the feedback guides reflection. Configure a reflection_lm (dspy.LM object) and select an auto mode ('light', 'medium', or 'heavy'). Pass these with your dspy.Module program and trainset to compile an optimized agent.

Inputs are a dspy.Module program, trainset (list of dspy.Example), metric (callable), reflection_lm, and auto mode. Output is a compiled_program (dspy.Module) optimized through reflective phases: defining feedback metrics, configuring the agent, and performing evolutionary search.

Yes. GEPA is production-ready for complex agents with tool use, improving reliability and performance through systematic reflective optimization. It requires no labeled data and works with instruction-only setups, making it practical for deployed workflows.