Four tools with basic descriptions and partial schemas. All tools have descriptions (10-200 chars, within baseline), but parameter schemas lack critical details: no enums for constrained inputs (e.g., 'mode' in query accepts 4 values but is free-form string), no min/max bounds, no format specifications. api_key parameter exposed in all tools violates secret-injection pattern. Error handling returns plain strings rather than structured guidance. Output schemas undocumented, LLMs cannot infer response structure. Tool names follow verb_noun convention (query, list_documents, get_entities, get_document) but descriptions are generic and lack WHEN/WHY context.
获取指定文档的原始 Markdown 内容。传入文档 ID(可从 list_documents 获取)。
获取知识图谱中的实体列表。可按文档筛选(传入 doc_id 参数)或查询全局实体。返回实体名称、类型、关联文档。
列出知识库中所有已索引的文档,包含标题、格式、导入日期和实体数量。
检索知识库。传入自然语言问题,返回相关文档段落、实体信息和关系描述。支持 hybrid(混合图检索,默认)、mix(混合图+原文块检索,推荐)、local(实体级)、global(关系级)四种模式。
API key exposed as tool parameter in all 4 tools. Credentials must use server-side secret injection via environment variables, not tool parameters. Agent traces log all parameters, secrets leak into logs and prompt history.
'mode' parameter in query tool accepts 4 specific values (hybrid, mix, local, global) but is defined as free-form string with no enum constraint. LLMs will hallucinate invalid modes. Enums are self-documenting and prevent invalid input.
No output schemas documented for any tool. Responses are plain strings (e.g., 'list_documents' returns markdown-formatted text). LLMs cannot parse structure or extract IDs for chaining downstream calls. Structured JSON responses with typed fields required.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 51 | 2026-07-28+ | v2 |
Error handling returns plain error strings (e.g., '查询失败: {e}') with no recovery guidance. LLMs cannot determine if error is retryable, user-fixable, or fatal. Errors must categorize the failure and suggest next steps.
No pagination support in list_documents or get_entities. If knowledge base grows large, returning all results will blow context window. Tools must accept limit/offset and return total count.