Single tool 'get_subtitles' with minimal definition quality. Tool has a reasonable name (verb_noun pattern) and a short description (68 chars), but critical gaps in parameter documentation and output schema definition. Input schema is present but parameter 'query' lacks a description field in the schema itself. No output schema is documented anywhere in the source code. No error handling guidance provided. The tool performs a semantic search operation but returns raw Milvus results without explaining the response structure to an LLM. No pagination, limits, or field documentation. This is typical of early-stage hobby projects, not production-ready tooling.

• Add a full output schema to the docstring. Document the response as: [{ season: int, episode: int, quote: str }], or whatever Milvus actually returns, so the LLM knows what fields to expect.
• Expand the 'query' parameter description to ~50-100 chars: 'Free-form text describing a Game of Thrones S1 scene, character, or quote. Examples: "Jon Snow's oath", "winter is coming", "Daenerys and dragons".' This guides LLM input.
• Add error handling documentation: 'Returns an empty list if no matches found. If the database is unavailable, returns an error message. If the query is too long (>1000 chars), truncates to the first 1000 characters.'
• Document result limits: 'Returns up to 5 most relevant quotes. To search again with different keywords, call the tool again with a new query.'
• Strip or document all response fields. If Milvus returns distance scores, clarify their meaning (e.g., 'higher distance = less relevant') or exclude them from the LLM response.
• Add a tool annotation (if supported by FastMCP): `readOnlyHint=true` to signal that this tool is safe to call without user confirmation.