Three read-only tools with clear naming (search_, fetch_, list_) and reasonable descriptions (100-150 chars). All tools have input schemas with types and descriptions. However, output schemas are not documented, LLMs cannot predict response structure for downstream chaining. Parameter descriptions are present but minimal (10-30 chars). No error handling guidance, no pagination limits documented, and no tool annotations (readOnlyHint). search_knowledge has good parameter coverage (query, top_k, mode, source_ids, document_ids) but fetch_document_chunks and list_sources lack depth. Missing: output field documentation, error recovery paths, and composition guidance.

• Document output schemas for all three tools. Example for search_knowledge: 'Returns array of {document_id: string, chunk_ordinal: int, text: string, score: float, source_id: string}'. This enables agents to extract and chain results.
• Add explicit pagination guidance: 'top_k accepts 1 - 100 (default 8). Results over 50 may exceed context limits; use source_ids or document_ids filters to narrow scope.'
• Expand parameter descriptions to 50 - 100 chars. For mode enum, add: 'hybrid (default) combines dense + sparse; dense uses embeddings; sparse uses keyword matching. Choose based on query type.'
• Add error recovery hints to tool descriptions. E.g., search_knowledge: 'If no results, try broader keywords or remove filters. Call list_sources to verify available documents.'
• Add tool annotations in MCP registration: readOnlyHint=true for all three tools. This signals to agents that these are safe to call without side effects.