Server has 5 tools with basic descriptions and input schemas visible in code. However, critical gaps undermine quality: (1) No output schemas documented anywhere, LLMs cannot predict return structure or chain tools; (2) Parameter descriptions are minimal (e.g., 'Collection name to ingest documents into' lacks format/constraint details); (3) No error handling guidance, tools fail silently without recovery hints; (4) Missing parameter validation rules (e.g., file_paths array size limits, top_k bounds); (5) No tool annotations (readOnlyHint, destructiveHint) despite clear risk levels. Naming is acceptable (verb_noun pattern), but descriptions average ~80 chars, below the 194-char baseline for A+ tools. Per-tool scores range 35 - 50, dragging the average down.

• Document output schema for each tool. E.g., search() returns {results: [{doc_id, text, score, metadata}], total_count, next_cursor}. Include field types and descriptions.
• Add bounds to numeric parameters: top_k (1 - 100, default 5), CHUNK_SIZE (64 - 2048), EMBED_BATCH_SIZE (1 - 32). Validate in code and return actionable errors.
• Implement confirmation pattern for delete_collection: add a 'confirm' boolean parameter or a separate 'confirm_delete_collection' tool that requires explicit approval.
• Enhance error messages with recovery hints. E.g., 'File not found: /path/to/file.pdf. Supported formats: .txt, .pdf, .md, .rst, .text. Check file path and retry.'
• Add tool annotations to schema: mark search, list_collections, get_collection_stats with readOnlyHint=true; mark delete_collection with destructiveHint=true; mark ingest_documents with idempotentHint=false.

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

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