feat: enhance search tool prompt for better relevance
Based on comprehensive research using 10 parallel agents analyzing:
- Exa search API best practices (neural search optimization)
- RAG query formulation patterns (HyDE, query decomposition)
- LLM hallucination prevention techniques
- Multi-step search strategies
- Search failure patterns in production systems
Key improvements:
- Query formulation: statement format with colon for neural search
(e.g., "Here is info about X:" vs "What is X?")
- Lost-in-the-middle mitigation: explicit instruction to read ENTIRE
result, not just beginning/end
- Grounding: attribution phrases ("According to search results...")
- Decomposition: clearer guidance for multi-part/comparison queries
- Uncertainty: explicit instruction to say "results don't contain X"
rather than guessing
- URL safety: consolidated citation guidance section
