Single tool 'retrieve_knowledge' has critical definition gaps. Input schema visible in code (Input struct with Query field) but lacks proper JSON Schema type declarations in the tool registration. Description is present (Chinese: '根据输入问题检索相似的知识点') but only 24 chars when translated, below the 34-char p10 baseline. No output schema documented in tool registration, only visible in code (Output struct). Parameter 'query' has a description in the struct tag but tool registration does not expose it. No error handling guidance, no pagination despite returning lists, no idempotent/destructive hints. Tool does one thing (retrieve) but implementation is incomplete for production use.

• Add JSON Schema to mcp.AddTool() registration: include 'properties' with 'query' type 'string', 'required' array, and 'description' for the parameter. Use mcp.Tool.InputSchema field.
• Document output schema in tool registration. Define a structured response with fields: query (string), knowledge_points (array of {id, content}), total_results (integer). Include in tool metadata or response documentation.
• Expand description to 50-150 chars: 'Retrieve similar knowledge points from the RAG index based on a search query. Returns up to 3 results ranked by relevance. Use this to find background information before answering user questions.'
• Add error handling with recovery guidance. Replace generic errors with: 'Query is empty. Provide a non-empty search query.' or 'Retrieval failed: [reason]. Retry with a simpler query or check if the knowledge base is initialized.'
• Add pagination support: accept optional 'limit' (1-50, default 10) and 'offset' (default 0) parameters. Return 'total_results' count and 'has_more' boolean so LLM can fetch additional pages.