JishuDB Knowledge Search / JishuDB知识库检索助手

Search existing JishuDB documents, list files, read text, and check import status; answer with source citations. No public-web research. / 查询JishuDB已有资料、文件清单、文档正文和导入进度，依据真实结果标注来源，不用于公网研究。

Retrieve facts from existing JishuDB documents and answer with source citations. Use the actual inventory, document, search, or job tools for each request.

先检索JishuDB已有资料，再依据真实结果回答并标注来源；清单、正文和任务进度使用对应工具。

At a glance

Item	Details
Use when	Questions concern documents, file inventory, or import status already in JishuDB.
Delivers	Answers with source filenames, verified inventories, document text, or observed job status.
Requires	An authorized JishuDB connection; the HTTP helper needs Bash and curl and only lists KBs or searches.
Does not	Import new material, research the public web, or infer file counts from search passages.

Use this skill when

• The user asks a factual question about material already stored in JishuDB.
• The user wants an exact knowledge-base or document inventory.
• The user asks to read a particular document or inspect its current metadata.
• The user asks whether an import has finished or uses /kb as a search request.

Do not use this skill when

Do not use this skill to import new material or research the public web. Use jishudb for setup or connection repair, research-report-hunter to discover external reports, and jishudb-plaud-import to archive selected Plaud material. The /kb field is a legacy hint, not a portable command registration; use the host's supported Skill loader if no such command exists.

Agent workflow

Route the request to the matching tool

• Knowledge-base names, IDs, or counts: call kb_list and use this run's results.
• Exact file names or counts: call kb_list_documents and follow every nextCursor. List only returned documents; search passages and job records are not file inventories.
• Current document metadata: identify the file with kb_list_documents, then call kb_get_document.
• Full document content: call kb_read_document_text; snippets do not reconstruct a full document.
• Semantic facts: call kb_search, synthesize returned passages, and cite sources.
• Import progress: call kb_list_jobs or kb_get_job. Jobs are processing attempts, so one document may have several jobs; do not equate job count with file count.
• If the needed tool fails or is unavailable, state what could not be verified.

The packaged kb-search.sh helper uses curl for the KB HTTP API. Set JISHUDB_URL for the actual backend; JishuShell panel bindings inject it automatically. The helper's default is http://host.docker.internal:18089. Set JISHUDB_TOKEN through the approved environment when authentication is enabled.

List knowledge bases when the target is unknown

These examples use repository-root paths. For a separately installed package, resolve kb-search.sh relative to the loaded Skill directory.

bash

bash skills/jishudb-search/kb-search.sh list

Use a relevant returned ID; the output includes names and chunk counts.

Search

bash

bash skills/jishudb-search/kb-search.sh search "question or keywords" <kb-id>

An optional third argument sets topK (default 6). When a panel binding supplies JISHUDB_DEFAULT_KB, the KB ID may be omitted:

bash

bash skills/jishudb-search/kb-search.sh "question or keywords"

Answer from retrieved evidence

Synthesize the returned passages and cite their actual source filenames. If the helper returns No matching passages found, report no matches for this search; try other keywords or a relevant KB from list without inventing evidence. Treat low-relevance passages cautiously and disclose uncertainty.

Errors and completion

The helper is read-only and can be called for several relevant queries. Keep a missing curl dependency distinct from a backend connectivity failure; verify the backend's actual running state and configured address. Complete the answer only to the extent supported by observed tool output.

Chinese reference / 中文操作参考

请求必须路由到对应工具

• 问“有哪些知识库”、知识库名称、标识或数量：调用 kb_list，只按本轮工具结果回答。
• 问某个知识库“有哪些文件”、精确文件名或文件数量：调用 kb_list_documents；若返回 nextCursor，继续翻页直到结束。只列返回的 documents，不得从检索片段、正文、历史对话或任务记录补出文件名。
• 问某个文件的当前状态或元数据：先用 kb_list_documents 确认文件，再调用 kb_get_document。
• 问某个文件具体写了什么：调用 kb_read_document_text；不要仅凭搜索摘要还原全文。
• 问资料中的语义事实：调用 kb_search，只依据返回片段综合，并标注来源。搜索结果不是完整文件清单。
• 问导入进度：调用 kb_list_jobs 或 kb_get_job。任务是处理尝试，同一文件可能对应多条任务，任务数不能当作文件数。
• 对应工具失败或不可用时，如实说明无法核实，不要根据记忆或字段名称猜测。

本技能目录里有一个零依赖脚本 kb-search.sh（用 curl 调 KB 的 HTTP API）。KB 地址由环境变量 JISHUDB_URL 配置（JishuShell 面板绑定时会自动注入；独立使用时自行 export，默认 http://host.docker.internal:18089）；若 KB 开了鉴权，另设 JISHUDB_TOKEN。

1. 不确定有哪些知识库时，先列出

bash

bash skills/jishudb-search/kb-search.sh list

输出每个库的 id / 名称 / 块数。挑一个相关的 id 用于检索。

2. 检索

bash

bash skills/jishudb-search/kb-search.sh search "用户的问题或关键词" <kb-id>

• 第 3 个参数可选 topK（默认 6）。
• 若面板绑定时设了默认库（JISHUDB_DEFAULT_KB），可省略 <kb-id>，直接：

  bash

  bash skills/jishudb-search/kb-search.sh "用户的问题"
  

返回形如:

text

Top 6 passages (cite the source filename in your answer):

[1] 01-quick-start.md  (relevance: high)
    JishuShell 是面向 Arm 边缘设备的 AI Agent 管理面板 …

[2] README.md  (relevance: medium)
    …

3. 据检索结果作答

• 综合这些片段写出答案，用自己的话组织。
• 标注来源：在答案里点名引用的文件名（如「据 01-quick-start.md…」），或在末尾列出来源。
• 若返回 No matching passages found，如实告知知识库里没有相关内容，不要编造；可建议换关键词或换一个库（先 list）。
• 相关度 low 的片段要谨慎使用，必要时说明置信度不高。

备注

• 脚本只读、无副作用，可安全多次调用（同一轮可对不同关键词检索几次再综合）。
• 若 kb-search: curl not found / 无法连接：说明运行环境缺 curl 或网络不通 KB，提示用户检查 KB 应用是否在面板里处于运行状态。

Resources

• kb-search.sh: Use the read-only HTTP helper for knowledge-base listing or passage search when its documented environment is available; other operations require host MCP tools.

Discovery

Keywords

• English: JishuDB search, /kb, passage retrieval, document inventory, ingest job status.
• Chinese: 查知识库、内部资料、文件清单、文档正文、导入进度.

Example requests

• “Search the existing knowledge base and cite the source filenames.”
• “Read this document's contents rather than reconstructing it from snippets.”
• “查知识库里的产品文档：支持哪些导入格式？标注来源。”
• “这个知识库有哪些文件？那份文档导入完成了吗？”

Nearby but different

• Application installation or connection repair → jishudb.
• Discover external research reports → research-report-hunter.