Single tool with strong naming and comprehensive schema, but descriptions lack LLM-specific optimization and lack error handling guidance. The tool name 'document_ingest_to_markdown' clearly conveys action, and input schema is well-defined with type constraints and defaults. However, descriptions are functional but brief, and output schema documentation is missing from the MCP tool registration. The tool definition is explicitly visible in src/mcp_rag/mcp/tools/document_ingest.py, so no inference penalty applies.

• Add output schema documentation to the MCP tool definition. Include field types and descriptions for markdown (string), metrics (object with pages_total, pages_processed, etc.), source_copy_path (string|null), and markdown_file_path (string|null). This enables LLMs to plan multi-step workflows.
• Expand tool description to include: (a) supported file types and max file size; (b) typical runtime ranges (e.g., '2 - 30 seconds for typical documents'); (c) error scenarios and recovery guidance (e.g., 'If vision API rate limit is hit, retry with vision_concurrency=1'); (d) prerequisites (e.g., 'GOOGLE_API_KEY must be set').
• Convert 'vision_model' parameter to an enum with explicitly supported options (e.g., ['gemini-2.5-flash', 'gemini-2.0-flash', 'gemini-1.5-pro']) rather than free-form string. Add description: 'Gemini model for semantic vision tasks. Use gemini-2.5-flash for speed, gemini-1.5-pro for accuracy.'
• Enhance 'file_path' parameter description to: 'Absolute path to document file. Supported formats: PDF, DOCX, PPTX, PNG, JPEG. Max 100 MB. Must be readable by the server process.'
• Add error classification guidance to tool description: 'Returns error if file not found (user should verify path), file too large (request smaller document), or vision API unavailable (retryable, contact support if persistent).'
• Document idempotency guarantee: 'This tool is idempotent, calling with the same file_path and parameters produces identical output (same markdown, same metrics). Safe to retry on transient failures.'

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

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