Four tools with basic verb-noun naming (list_, create_, delete_, get_) but significant quality gaps. All tools have descriptions (10-60 chars), but most are too brief to guide LLM selection. Input schemas present for 3/4 tools but lack parameter descriptions entirely. No output schemas documented. No error handling guidance. No tool annotations (readOnlyHint, destructiveHint). Destructive tool (delete_database) has no confirmation or dry-run pattern. Factory pattern tool (get_database) description is vague about auto-detection behavior.

• Add 50-150 character descriptions to each tool explaining WHAT it does, WHEN to use it, and WHAT it returns. Example: 'List all available vector databases with their metadata schemas. Call this first to discover existing databases before creating or connecting to one.'
• Document output schemas for all tools. Specify fields returned (e.g., database_id, name, created_at, metadata_schema) with types so LLMs can chain calls.
• Add descriptions to all input parameters. For 'name': 'Database name (alphanumeric, 1-255 chars, must be unique)'. For 'metadata_schema': 'Optional JSON Schema object defining custom metadata fields for documents.'
• Implement confirmation pattern for delete_database: add optional 'confirm=true' parameter and return 'This will permanently delete the database. Call again with confirm=true to proceed.' on first call.
• Add tool annotations: mark list_databases and get_database with readOnlyHint=true; mark create_database and delete_database with destructiveHint=true (delete only).

Inferred effective spec: <=2025-11-25.

• Relies on Logging (deprecated) - log to stderr or use OpenTelemetry