Server has 5 tools with basic descriptions and input schemas. Naming follows verb_noun convention (list_, search_, extract_, get_, health_check). However, descriptions are generic and lack actionable context for LLM selection. Parameters have types but descriptions are minimal. Output schemas are not documented, tools return unstructured strings rather than typed objects. Error handling returns raw error messages without recovery guidance. No tool annotations (readOnlyHint, etc.). Composition is reasonable but output formatting is suboptimal for agent reasoning.

• Convert all tool responses from strings to typed JSON objects. Example: extract_all should return {results: [{date: string, score: float, content: string}], total_count: int, query: string} instead of formatted text.
• Add explicit parameter constraints: top_k should have minimum=1, maximum=100; query should have minLength=1, maxLength=500. Document these in parameter descriptions.
• Enhance tool descriptions with WHEN/WHY context. Example: 'Use extract_all to find historical documents similar to a query within a specific collection. Use search_web for current information not in your database.'
• Implement structured error responses: {error: string, code: string, retryable: bool, suggestion: string}. Example: {error: 'Collection not found', code: 'COLLECTION_NOT_FOUND', retryable: false, suggestion: 'Call list_collections() to see available collections.'}
• Add tool annotations to schema: mark all tools with readOnlyHint=true since they do not modify state. This enables agents to reason about safety.