Single tool with clear purpose and well-structured schema. Tool name 'vslm_predict' uses verb_noun pattern appropriately. Description is comprehensive (287 chars, within 10-1024 baseline) and explains WHAT, WHEN, and WHY. Input schema is complete with proper types and descriptions. However, output schema is documented only in docstring, not in formal schema definition. Error handling is present but minimal, no recovery guidance for common failures (auth, quota, upstream). No tool annotations (readOnlyHint, idempotentHint) despite being read-only and idempotent. Missing pagination guidance despite accepting ~2000 sentences per call.

• Add formal output schema to @mcp.tool() decorator: specify relevant_sentences (array of strings), total_sentences (int), relevant_count (int), topic_used (string), topic_source (enum: 'zettaquant'|'caller'), model_id (string|null).
• Add tool annotations: @mcp.tool(readOnlyHint=True, idempotentHint=True) to signal safety properties to agents.
• Enhance error messages with recovery guidance: 'ZettaQuant quota exceeded (429). Quota resets at {resets_at}. Retry after reset or contact support.' for 429; 'ZettaQuant auth failed (401). Check ZQ_API_KEY in MCP host config and restart.' for 401.
• Document batch size limits in description: 'Accepts up to 2000 sentences per call. For larger datasets, split into multiple calls and merge results.'
• Return explicit error for empty input instead of silent early return: raise ValueError('sentences list cannot be empty') with clear message.