Critical gaps across all dimensions. Tool definitions lack proper schemas, no JSON Schema visible in source code. Descriptions are present but generic (avg ~80 chars, below 194-char baseline). Parameters lack type definitions and constraints. No error handling guidance, no output schema documentation, no security considerations. Tool registry exists but does not expose structured MCP-compliant definitions. This is a custom FastAPI wrapper around LangChain, not a proper MCP server implementation.

• Add formal JSON Schema definitions for all tool inputs. Each parameter must have 'type', 'description' (50 - 150 chars), and constraints (enum, minLength, pattern, etc.). Example: {"image_data": {"type": "string", "format": "base64", "description": "Base64-encoded image (JPEG/PNG, max 10MB). Used for medical image analysis via CLIP embeddings."}}
• Expand tool descriptions to 150 - 250 chars. Include: (1) What it does, (2) When to use it vs similar tools, (3) What it returns. Example: 'Analyze medical images using CLIP embeddings and vision models. Call this when the user provides an image of a prescription, scan, or medical document. Returns extracted text, detected entities (drug names, dosages), and confidence scores.'
• Document output schemas for each tool. Specify return type (object/array), required fields, and field types. Example: analyze_image returns {"extracted_text": string, "entities": [{"type": string, "value": string, "confidence": number}], "analysis_time_ms": number}
• Add error handling guidance. Document common failures (e.g., 'Unsupported image format', 'PDF too large', 'API rate limit exceeded') and recovery steps. Example: 'If image analysis fails with "unsupported format", try converting to JPEG or PNG first. If rate-limited, retry after 60 seconds.'
• Implement parameter validation with clear error messages. When web_search receives num_results > 100, return 'Invalid num_results: 150. Must be 1 - 100.' instead of a generic 500 error. This enables LLM self-correction.