AxGEPA GEPA

Use the generated AxGEPA optimizer surface for portable GEPA tuning. TypeScript also exposes the convenience optimize() helper; generated packages currently keep the optimizer explicit.

engine := axllm.NewAxGEPA(reflectionClient, nil)
result := engine.Optimize(request, evaluator)

The optimizer runs a program against examples, scores predictions, searches for better instructions/demos/component settings, and returns an artifact that can be applied immediately or serialized.

What It Does

GEPA explores changes and keeps tradeoffs visible through Pareto-aware results. The convenience helper runs bootstrap demos for small starter sets before GEPA, which gives the optimizer concrete examples to work with.

flowchart LR
  A["Program"] --> D["GEPA"]
  B["Train examples"] --> D
  C["Metric or judge"] --> D
  D --> E["Reflective mutation"]
  E --> F["Pareto evaluation"]
  F -->|explore| E
  F --> G["Optimized artifact"]
  G --> H["Save / load / apply"]

Core Call Shape

result = optimize(program, examples, metric, options)
program.applyOptimization(result.optimizedProgram)

Generated language packages expose the optimizer surface available in their AxIR API, usually as an AxGEPA engine.

Baseline Reporting In TypeScript

TypeScript optimize(...) and AxGEPA.compile(), and generated-language GEPA engines and helpers, return baselineScore and baselineScores for the program before bootstrap or prompt changes. Generated artifacts also retain these fields in metadata, alongside bestScore. Existing caller-supplied demos are part of that baseline. Compare bestScore - baselineScore to report the change, even when the original candidate is no longer on the Pareto frontier.

Both scores use the same validation subset, sampling and metric or adapter. baselineScores holds averaged objective values ({ score: value } for scalar metrics), and baselineScore applies the same scalarization as bestScore. These fields are optional on the shared optimizer result because other optimizers may omit them. Save the baseline alongside the optimized artifact when recording experiments.

When bootstrap adds demos, allow two validation passes within maxMetricCalls; without new demos, GEPA reuses the initial evaluation. The helper’s implicit budget is at least 100 and grows when needed to fit those passes; an explicit budget stays unchanged. Bootstrap demo generation retains its separate limits and collects fresh training rows, excluding stale held-out traces from inactive children. Generated artifacts report validation/search calls as metadata.totalMetricCalls, bootstrap calls as metadata.bootstrapMetricCalls, and their sum as metadata.totalCalls.

For Rust, use optimize_with_client(&mut program, Rc<RefCell<Client>>, examples, options) to evaluate a live student. The client-free optimize(...) helper requires explicitly pre-scored tasks. Supply an evaluator to AxGEPA when scoring predictions with an application-specific metric; the public GEPA examples demonstrate this in every language.

Common Patterns

• Start with deterministic eval fixtures for the metric.
• Keep maxMetricCalls explicit.
• Use scalar metrics for one clear objective.
• Use multi-objective metrics for tradeoffs such as accuracy/cost or quality/brevity.
• Keep validation examples separate when the optimizer surface supports them.
• Save artifacts with enough metadata to explain the examples and metric that produced them.

AxGen

gepa := ax.NewAxGEPA(map[string]ax.Value{"maxMetricCalls": 40})
artifact, err := gepa.Optimize(request, evaluator)

Flow

gepa := ax.NewAxGEPA(map[string]ax.Value{"maxMetricCalls": 40})
artifact, err := gepa.Optimize(request, evaluator)

Agent

gepa := ax.NewAxGEPA(map[string]ax.Value{"maxMetricCalls": 40})
artifact, err := gepa.Optimize(request, evaluator)

Artifact

gepa := ax.NewAxGEPA(map[string]ax.Value{"maxMetricCalls": 40})
artifact, err := gepa.Optimize(request, evaluator)

Production Notes

Optimization is a build-time or evaluation-time workflow, not something to hide in every request path. Track score history, cost, token usage, selected Pareto point, and the final artifact version.

See optimize() API.