config_adaevolve.yaml

# AdaEvolve Configuration for HotPotQA Prompt Optimization
# Evolves a natural-language instruction prompt (not code) for multi-hop QA
#
# Usage:
#   uv run skydiscover-run initial_prompt.txt evaluator.py -c config_adaevolve.yaml --search adaevolve

max_iterations: 100
checkpoint_interval: 1
log_level: "INFO"
diff_based_generation: false  # Full rewrite mode (essential for prompt optimization — can't diff free text)
max_solution_length: 10000
language: "text"  # Tells framework this is plain text, not code

llm:
  models:
    - name: "gpt-5-mini"
      weight: 1.0
  api_base: "https://api.openai.com/v1"
  temperature: 0.7
  max_tokens: 16000
  timeout: 300

search:
  type: adaevolve
  num_context_programs: 4
  database:
    population_size: 20
    num_islands: 2
    migration_interval: 10
    migration_count: 5
    decay: 0.9
    intensity_min: 0.15
    intensity_max: 0.5
    use_unified_archive: true
    fitness_weight: 1.0
    novelty_weight: 0.0
    diversity_strategy: text
    use_adaptive_search: true
    use_ucb_selection: true
    use_migration: true
    use_dynamic_islands: true
    max_islands: 5
    spawn_productivity_threshold: 0.015
    spawn_cooldown_iterations: 30
    use_paradigm_breakthrough: true
    paradigm_window_size: 10
    paradigm_improvement_threshold: 0.12
    paradigm_max_uses: 2
    paradigm_max_tried: 10
    paradigm_num_to_generate: 3
    enable_error_retry: true
    max_error_retries: 2

prompt:
  system_message: |
    You are an expert prompt engineer. Your task is to revise an group of existing prompts designed for large language models (LLMs), without being explicitly told what the task is.

    Your improvements should:

    * Infer the intended task and expected output format based on the structure and language of the original prompts.
    * Clarify vague instructions, eliminate ambiguity, and improve overall interpretability for the LLM.
    * Strengthen alignment between the prompt and the desired task outcome, ensuring more consistent and accurate responses.
    * Improve robustness against edge cases or unclear input phrasing.
    * If helpful, include formatting instructions, boundary conditions, or illustrative examples that reinforce the LLM's expected behavior.
    * Avoid adding unnecessary verbosity or assumptions not grounded in the original prompt.
    * Use feedback to help reduce known errors

    The revised prompt should maintain the same input interface but be more effective, reliable, and production-ready for LLM use.

evaluator:
  timeout: 1800  # 30 min — each eval runs ~300 LLM calls with BM25 retrieval
  max_retries: 3
  cascade_evaluation: false