Chapter 5

GEPA: Genetic-Pareto Optimization

Multi-objective prompt optimization using genetic algorithms and Pareto fronts.

Key Concepts

GEPA merges genetic algorithms (selection, crossover, mutation, elitism) with Pareto optimization (multi-objective trade-offs). Its distinguishing feature is the use of natural language reflections to guide the evolutionary process.

Basic Usage

from gepa import GEPAOptimizer

optimizer = GEPAOptimizer(
    population_size=20,
    generations=10,
    mutation_rate=0.2,
    crossover_rate=0.7,
    objectives=["accuracy", "clarity", "completeness"],
    reflection_model="gpt-4"
)

compiled = optimizer.compile(
    program=program,
    trainset=training_data,
    valset=validation_data
)
Next: State-Space Optimization