Loading ToolBench
Baichuan medical AI model MCP service integration supporting Baichuan-M3-Plus and Baichuan-M2-Plus models with medical Q&A, file upload, and document Q&A capabilities
5 tools with mixed quality. Tool names follow verb_noun convention (baixiaoying_chat, baixiaoying_upload_file, etc.), but descriptions are in Chinese and vary in clarity. Input schemas are present and typed for all tools, with enums used appropriately (e.g., model selection, evidence_scope). However, parameter descriptions are minimal (10-30 chars typically), and output schemas are not documented in the source. Error handling and recovery guidance are absent. The server exposes resource management (file upload/delete) but lacks proper destructive operation safeguards (no dry-run, no confirmation pattern). Medical domain requires high precision, current definitions do not meet production safety bars.
使用百小应大模型进行医学问答对话，支持 Baichuan-M3-Plus 和 Baichuan-M2-Plus 模型，可处理文本和医学文档
删除指定的已上传文件。被知识库使用的文件需要先解除关联后才能删除。
查询指定文件的解析状态。文件需要解析完成（状态为 online）后才能用于对话。
获取已上传的文件列表
上传医学文档用于后续的文档问答。支持 pdf、doc、docx、txt、html、md、csv、png、jpg 等格式。
Output schemas not documented. LLM cannot infer what fields to expect from tool responses (e.g., what fields does list_files return? What are possible file_status values?). Forces trial-and-error parsing of responses.
Parameter descriptions are terse (10-30 chars) and lack format/constraint guidance. E.g., file_path has no length limit documented, temperature range is in schema but not explained in description. Descriptions should state expected format, range, and constraints explicitly.
No error recovery guidance. Tools like delete_file and upload_file have preconditions and failure modes (file in use, unsupported format, size limit) but descriptions do not explain what the LLM should do if the operation fails. Missing 'If X fails, try Y' recovery hints.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | D | 59 | <=2025-11-25 | v2 |
| 2026-03-09 | F | 49 | - | v1 |
Destructive operations (delete_file) lack confirmation or dry-run pattern. Medical domain demands reversibility checks or at least explicit confirmation before irreversible deletion. Current design permits silent deletion.
Tool descriptions in Chinese may not be parsed by all LLMs equally. While valid, descriptions should provide both language versions or prioritize English for maximum agent compatibility. Current descriptions are domain-specific and may not clearly signal intent to agents unfamiliar with medical terminology.
File type support listed in description ('pdf, doc, docx, txt, html, md, csv, png, jpg') but no explicit enum constraint in schema. LLM may attempt to upload unsupported formats. Should validate against whitelist and return clear error.
No pagination documented for list_files. If many files exist, response may blow context window. Common pattern: list tools should accept limit/offset and return total_count.