Loading ToolBench
MCP Server implementation on Cloudflare Workers with Streamable HTTP Transport for RAG system built on Cloudflare Vectorize
Single tool 'search' has a complete JSON Schema with proper typing and constraints, and a non-empty description in Japanese. However, the description is brief (48 chars in English equivalent), lacks English localization making it inaccessible to many agents, provides no guidance on when to use it or what it returns, and the tool implementation is incomplete (mock response only). No output schema is documented. The tool exhibits acceptable baseline compliance but significant gaps in LLM-optimization patterns. Most production agents would struggle with the Japanese-only description and lack of recovery guidance.
Cloudflare Vectorizeに構築されたRAGシステムを検索
Tool description is in Japanese only ('Cloudflareに構築されたRAGシステムを検索') with no English translation. LLMs operating in English-dominant environments will be unable to parse this description, leading to incorrect tool selection or avoidance.
Description is only 48 characters (below the 10-1024 ideal range of production tools and well below the recommended 50-200 chars for LLM optimization). It lacks WHAT the tool does, WHEN to use it, and WHAT it returns. Agents will struggle to decide when this tool is appropriate vs alternatives.
No output schema documented. The code shows a mock response structure with 'results', 'totalCount', 'processingTime' but the tool definition provides no schema documentation for LLMs to reason about downstream tool chaining or result field extraction.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 49 | <=2025-11-25 | v2 |
| 2026-03-09 | F | 45 | 2024-11-05+ | v1 |
Tool implementation returns mock data only ('TODO: 実際の検索処理を実装（次のタスクで実装）'). The tool is non-functional for production use, returning hardcoded sample results regardless of input query. This violates the principle that tools must produce contextually appropriate responses.
Parameter descriptions lack guidance on valid input formats and constraints. For example, 'query' description is just '検索クエリ（自然言語）' with no mention of minimum length, character restrictions, or examples of well-formed queries.
No error handling guidance. The tool schema provides no information about failure modes (e.g., invalid index name, network failures, threshold filtering too strict) or recovery paths for LLMs when search fails.
Default value for 'indexName' is 'rag-index' but description does not explain what happens if this index does not exist on the Cloudflare Vectorize account, or how to discover available indices.
No documented result limits. The tool accepts topK up to 100, but does not explain truncation behavior or pagination strategy if more results are available. Large result sets risk context window exhaustion.