A multi-modal voice assistant with semantic routing to specialized MCP tools (RAG, web search, structured data lookup). Integrates ASR (Deepgram), LLM (Claude), TTS (ElevenLabs), and LangGraph-based agent orchestration.
Three tools with minimal descriptions and incomplete schemas. lookup_structured and ask_documents have basic parameter descriptions but lack output schema documentation. search_web has no parameter description at all (query param is undescribed). All three tools lack error handling guidance, recovery instructions, and actionable error messages. No tool annotations (readOnlyHint, etc.). Descriptions are functional but generic (avg ~120 chars), below the 194-char baseline for A+ tools. No pagination support despite search_web returning lists. Parameter validation rules not documented.
Query the internal knowledge base. Use for questions about company policy, product docs, technical guides, or any domain knowledge stored in indexed documents.
Look up specific structured facts. Available keys: latency_budget, tech_stack, supported_languages, routing_method, web_search_provider
Search the web for current information, recent news, or facts not in the document knowledge base. Use for anything requiring up-to-date information.
search_web query parameter has no description. LLMs cannot infer whether to pass natural language, keywords, or structured queries.
No output schemas documented for any tool. LLMs cannot plan downstream calls or extract required fields (e.g., does search_web return URLs, titles, snippets, or all three?).
search_web returns raw list with no pagination, limit, or result count. Unbounded results risk context window exhaustion.
No error handling guidance. ask_documents calls external HTTP service (localhost:8000) with 10s timeout but provides no recovery instructions if service is down.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 54 | 2026-07-28+ | v2 |
lookup_structured returns free-text values (e.g., 'ASR 300ms + rewrite 65ms...') instead of structured objects. LLMs must parse unstructured text, wasting tokens and risking misinterpretation.