Module optimizer 

Available on crate feature eval only.

Prompt optimization engine.

Iteratively improves an agent’s system instructions using an optimizer LLM and an evaluation set. Used by the adk optimize CLI command.

Overview

The PromptOptimizer runs an optimization loop:

1. Evaluate the agent against the eval set to get a baseline score
2. If the score already meets the target threshold, report “no optimization needed”
3. Otherwise, ask the optimizer LLM to propose improved instructions
4. Apply the best improvement and re-evaluate
5. Repeat until max iterations or target threshold is reached
6. Write the best-performing instructions to the output file

Example

use adk_eval::optimizer::{PromptOptimizer, OptimizerConfig};
use std::sync::Arc;

let optimizer = PromptOptimizer::new(
    optimizer_llm,
    evaluator,
    OptimizerConfig::default(),
);
let result = optimizer.optimize(agent, &eval_set).await?;
println!("Final score: {}", result.final_score);

Structs

OptimizationResult

Result of a prompt optimization run.

OptimizerConfig

Configuration for the prompt optimization loop.

PromptOptimizer

Iteratively improves an agent’s system instructions using an optimizer LLM and an evaluation set.

Functions

run_optimization_loop

Run the core optimization loop with injectable evaluation and proposal functions.