mcp-qdrant provides 7 read-only semantic search tools with HTTP transport and basic Qdrant integration. All tools have descriptions (10-50 chars) and schemas with typed parameters, placing it in the Fair-to-Good range. However, descriptions are very brief and lack context for when/why to use each tool. Parameters are documented but lack guidance on constraints, ranges, or expected formats. No tool annotations (readOnlyHint/destructiveHint) are present. Error handling is not visible in the source. Output schemas are not explicitly documented in the tool definitions. The tool names follow verb_noun convention (search_*, filter_*, count_*, get_*), which is correct, but some tools have overlapping functionality (search_text vs filter_search vs keyword_search) without clear differentiation in descriptions. No evidence of pagination guidance in scroll_points despite being the pagination tool. Overall, definitions are present and minimally functional but fall short of production-grade LLM-optimization patterns.

• Expand all tool descriptions to 150 - 250 characters. Include: WHAT it does, WHEN to use it vs similar tools, and WHAT it returns. Example: 'search_text: Perform semantic search by converting your query to embeddings and finding the most similar stored vectors. Use this for open-ended semantic queries. Returns top N matching points with IDs, scores, and optional payloads. Faster than filter_search when no metadata filtering needed.'
• Clearly differentiate search_text, filter_search, and keyword_search in descriptions: (1) search_text = pure semantic similarity, (2) filter_search = semantic + metadata filtering (use when you need to narrow by username/filename/etc), (3) keyword_search = semantic + keyword substring matching. Mention use-cases for each.
• Add explicit output schemas to all tools. Document what fields are returned, their types, and whether they're always present. Example for search_text: 'Returns { results: [{ id: string, score: float, text: string, payload?: object }], collection_name: string }'.
• Add readOnlyHint: true to all 7 tools in the tool registration (source code). This tells agents these calls are safe to retry and have no side effects.
• Document pagination for scroll_points: 'Paginate through all points in the collection. Returns a maximum of `limit` points per call. Use the returned `next_offset` in the next call to fetch the next page. When `next_offset` is null, you've reached the end.'
• Add constraint guidance to numeric parameters: 'limit: integer, 1 - 1000, default 10. Larger limits increase latency. For initial exploration, use 10 - 20. For batch operations, use up to 100.'

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

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