A RAG system with MCP integration that crawls sitemaps, chunks content semantically, embeds text, and provides context retrieval via MCP tools
Single tool 'get_context' has a verb-noun name (positive), but the description is vague and under 20 chars ('Get relevant tech context to the query.'). Input schema is present with 'url' parameter, but lacks type validation details, format constraints, or error guidance. No output schema is documented, the tool returns a dict from crawl_sequentially() but the structure is never specified. No parameter description for 'url' beyond the schema hint. No error handling guidance. The tool performs a complex RAG pipeline (crawl, chunk, embed, store) but exposes only a single parameter, hiding critical complexity from the LLM. Missing descriptions for what the output contains, when to use this tool, or how to handle failures.
Get relevant tech context to the query.
Tool description is under 20 characters and provides no actionable context. 'Get relevant tech context to the query.' does not explain what 'context' means, when to call this tool, what it returns, or how to handle errors.
No output schema documented. The tool returns a dict with 'success', 'error', 'total_urls', 'successful', 'failed', 'total_content_length', 'processed_urls' fields (inferred from crawl.py), but LLMs cannot plan downstream actions without knowing the response structure.
Parameter 'url' lacks a description in the tool definition. The schema hint says 'a sitemap url (ex: https://modelcontextprotocol.io/sitemap.xml)' but this is an example, not a constraint. No format validation, no guidance on what happens if the URL is invalid or unreachable.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 38 | 2026-07-28+ | v2 |
No error handling guidance. The tool can fail if the sitemap is malformed, URLs are unreachable, crawling times out, or embeddings fail. No recovery hints for the LLM (e.g., 'If crawling fails, try a smaller sitemap or check the URL format').
Tool performs multiple responsibilities (crawl, chunk, embed, store in Qdrant) but exposes only one parameter. The LLM cannot control chunking strategy, embedding model, or storage behavior. This violates single-responsibility principle and limits composability.