Configurable Solution Generation Prompts

[alirezamshi]

Description

The solution_generation method has a hardcoded prompt. Users can only configure LLM models but cannot customize the generation behavior or provide domain-specific guidance.

Current limitation:

# llm_tree_search.py L433-440
solution_generation_prompt = """
    You are a world class problem solver. Generate {per_step_solution_count} exact solutions...
"""

Goal

Allow users to customize solution generation via config:

• Custom system prompt template
• Domain-specific directions/constraints
• Solution format requirements

Proposed Config

search_strategy:
  type: "llm_tree_search"
  params:
    idea_generation_model: "gpt-4.1"
    idea_generation_prompt: |
      You are an expert in {domain}. Generate solutions that...
    idea_generation_directions:
      - "Prefer memory-efficient approaches"
      - "Consider edge cases for large inputs"
    idea_generation_format: "numbered_list"  # or "xml_tags"

Implementation

1. Add new params to LlmSteeredTreeSearch.__init__
2. Template the solution_generation_prompt with user values
3. Update strategies.yaml with new param defaults

References

• solution_generation method
• src/execution/search_strategies/strategies.yaml - Strategy presets
• src/config.yaml - Mode configurations