koog-agents-optimization/ai.koog.agents.optimization.optimizers.gepa

Package-level declarations

Types

GEPAFeedback

fun interface GEPAFeedback<Input, Output, InputLabel>

Converts an evaluated rollout into the continuous score and textual module feedback GEPA needs.

GEPAFeedbackResult

data class GEPAFeedbackResult(val score: Double, val moduleFeedbacks: Map<String, GEPAModuleFeedbackResult?>)

Feedback produced for one GEPA rollout.

GEPAMergeConfig

data class GEPAMergeConfig(val enabled: Boolean = false, val maxMergeInvocations: Int = 5, val maxMergeAttempts: Int = 10, val maxFindSharedAncestorAttempts: Int = 10, val valSetSubsampleSize: Int = 5)

Runtime controls for GEPA's optional merge step.

GEPAModuleFeedbackResult

data class GEPAModuleFeedbackResult(val input: String, val output: String, val feedbackText: String)

Textual feedback for one optimizable module on one GEPA rollout.

GEPAModuleSelectionStrategy

enum GEPAModuleSelectionStrategy : Enum<GEPAModuleSelectionStrategy>

Strategy for choosing which optimizable modules GEPA updates on each feedback batch.

GEPAOptimizer

class GEPAOptimizer<Input, Output, InputLabel>(gepaFeedback: GEPAFeedback<Input, Output, InputLabel>, feedbackValSplitFn: (TrainSet<Input, InputLabel>) -> GEPATrainSetSplit<Input, InputLabel>, reflectionModel: LLModel, val storagePath: ResilientPath, numRollouts: Int, feedbackBatchSize: Int = 3, skipPerfectFeedbackBatches: Boolean = true, perfectScoreThreshold: Double = 1.0, randomSeed: Long = 42, moduleSelectionStrategy: GEPAModuleSelectionStrategy = GEPAModuleSelectionStrategy.ROUND_ROBIN, useStructuredOutput: Boolean = false, requireThinkingFieldInOutput: Boolean = false, mergeConfig: GEPAMergeConfig = GEPAMergeConfig(), failureScore: Double = 0.0, feedbackFailureRateThreshold: Double = 1.0, validationFailureRateThreshold: Double = 0.9, seedValidationFailureRateThreshold: Double = 0.0, abortOnFailureRateExceeded: Boolean = true) : AgentOptimizer<Input, Output, InputLabel>

GEPA optimizer, based on "GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning" (https://arxiv.org/abs/2507.19457).

GEPATrainSetSplit

data class GEPATrainSetSplit<Input, InputLabel>(val feedbackSet: TrainSet<Input, InputLabel>, val validationSet: TrainSet<Input, InputLabel>)

The two sets GEPA trains on, split off a training set.

Functions

singleModuleGepaFeedback

fun <Input, Output, InputLabel> singleModuleGepaFeedback(moduleName: String, formatInput: (Input) -> String = { it.toString() }, formatOutput: (Output) -> String = { it.toString() }, feedback: (input: Input, output: Output, gold: InputLabel, fullTrace: Prompt?, subgraphTraces: Map<String, List<Demonstration>>) -> Pair<Double, String>): GEPAFeedback<Input, Output, InputLabel>

Convenience helper for building a feedback function for a single explicit module.

strategyOnlyGepaFeedback

fun <Input, Output, InputLabel> strategyOnlyGepaFeedback(formatInput: (Input) -> String = { it.toString() }, formatOutput: (Output) -> String = { it.toString() }, feedback: (input: Input, output: Output, gold: InputLabel, fullTrace: Prompt?, subgraphTraces: Map<String, List<Demonstration>>) -> Pair<Double, String>): GEPAFeedback<Input, Output, InputLabel>

Convenience helper for building a feedback function for agents that only optimize the strategy instruction.