Single tool 'search_zoning' has a clear action verb and reasonable description (95 chars), but critical gaps undermine quality. Input schema is present with proper type and description for the 'query' parameter. However, NO output schema is documented, the description mentions 'structured citations with filename, section, and relevant text snippets' but the actual response structure is not formally defined. This forces LLMs to guess at output fields and breaks downstream tool chaining. Error handling is absent from the code sample. The tool lacks idempotency hints, pagination support (if results are large), and guidance on when to use it vs. alternatives. No tool annotations (readOnlyHint, etc.) are present. The server exposes OpenAI API integration details (vector store management) but does not abstract them into tool parameters, violating the chat-data-model principle.

• Document the complete output schema: define the structure of 'citations' (array of objects with 'filename', 'section', 'text' fields, plus any metadata like relevance_score or page_number). Include example output.
• Add error handling responses: specify what happens if no results match, if the vector store is unavailable, or if the query is malformed. Provide recovery guidance (e.g., 'Try a simpler query' or 'Check vector store status').
• Implement pagination: add optional 'limit' (default 20, max 100) and 'offset' parameters. Return total_count and next_offset in the response so agents can iterate through large result sets.
• Add tool annotations: include 'readOnlyHint: true' and 'idempotentHint: true' in the tool definition so LLMs know this is safe to retry and has no side effects.
• Enhance parameter description: specify query length (e.g., 5 - 500 characters), mention that queries are semantic (not regex), and provide an example: 'setback requirements for residential buildings'.