GEPA

// Overview

GEPA (Genetic-Pareto) is a Python framework that improves any system with textual parameters against an evaluation metric you define. Instead of collapsing a run into a single number, it uses a language model to read full execution traces (error messages, profiling data, reasoning logs) to work out why a candidate failed, then proposes targeted fixes through iterative reflection, mutation, and Pareto-aware selection.

It is built for developers who tune prompts and AI pipelines and want a systematic way to search for better variants with relatively few evaluations. The same approach extends beyond prompts to code, agent architectures, configurations, and other text artifacts through the optimize_anything API.

As a prompt-programming and LLM-orchestration tool, GEPA fits alongside DSPy, where it is available as dspy.GEPA for optimizing AI programs. You can use it standalone with its own gepa.optimize entry point or inside a DSPy pipeline.

// What it does

• Reflective optimization: an LLM reads full execution traces to diagnose failures rather than relying on a single scalar reward
• Pareto-aware evolutionary search that keeps candidates excelling on different task subsets and evolves new variants from them
• Reaches gains in roughly 100-500 evaluations, far fewer than reinforcement-learning approaches like GRPO
• optimize_anything API to tune any text artifact: prompts, code, agent architectures, configs, or SVGs
• Integrates with DSPy as dspy.GEPA for optimizing AI pipelines
• Configurable task and reflection language models, so you can pair a cheaper task model with a stronger reflection model

// Getting started

Install from PyPI, then optimize a seed prompt against a dataset with a metric you choose.

pip install gepa

import gepa

trainset, valset, _ = gepa.examples.aime.init_dataset()

seed_prompt = {
    "system_prompt": "You are a helpful assistant. Answer the question. "
                     "Put your final answer in the format '### <answer>'"
}

result = gepa.optimize(
    seed_candidate=seed_prompt,
    trainset=trainset,
    valset=valset,
    task_lm="openai/gpt-4.1-mini",
    max_metric_calls=150,
    reflection_lm="openai/gpt-5",
)

print("Optimized prompt:", result.best_candidate['system_prompt'])

import dspy

optimizer = dspy.GEPA(
    metric=your_metric,
    max_metric_calls=150,
    reflection_lm="openai/gpt-5",
)
optimized_program = optimizer.compile(student=MyProgram(), trainset=trainset, valset=valset)

Commands and code are distilled from the project's own documentation — always check the official repo for the latest.

// When to use it

• Improve a system or task prompt against a benchmark or custom metric without hand-tuning by trial and error
• Optimize a DSPy AI program end to end using dspy.GEPA
• Tune non-prompt text artifacts such as code, agent architectures, or configuration files with the optimize_anything API
• Search for better-performing variants when reinforcement-learning approaches would need too many evaluations

// How GEPA compares

GEPA alongside other open-source prompt programming tools AI/TLDR tracks, ranked by GitHub stars.