Agentic RAG e-commerce assistant over scraped product reviews, with MCP-served retrieval and web search
Two tools with minimal definitions. Both have verb-noun naming (get_product_info, web_search) and single-string input schemas, but descriptions are generic (avg 85 chars, below 194-char baseline). No parameter descriptions provided. No output schemas documented. No error handling guidance beyond generic exception messages. Tool composition is reasonable (separate concerns), but schema completeness and description depth fall well short of production standards.
Retrieve product information for a given query from local retriever.
Search the web using DuckDuckGo if retriever has no results.
Parameter descriptions missing. Both tools accept 'query' string but provide no guidance on format, length, or expected content. LLMs cannot infer whether to pass a product name, category, or full sentence.
Output schemas not documented. Tools return unstructured strings (e.g., 'No local results found' or raw DuckDuckGo output). LLMs cannot parse structured fields or plan downstream calls. Baseline: 100% of A+ tools document return types.
Tool descriptions are generic and lack context. 'Retrieve product information for a given query from local retriever' does not explain when to use this vs web_search, what metadata is returned, or what 'local retriever' means to the LLM.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 51 | <=2025-11-25 | v2 |
Error handling provides no recovery guidance. Both tools catch exceptions and return generic error strings (e.g., 'Error retrieving product info: ...'). LLMs cannot determine if the error is retryable, user-fixable, or fatal.
No pagination or result limits documented. get_product_info returns all retriever results as a single formatted string; web_search returns raw DuckDuckGo output. Large result sets will blow context windows. Baseline: tools returning lists should accept limit/offset and return counts.