dsh-workspace-api · Enterprise Knowledge Q&A Agent / 企业信息查询 Agent

> Turn your enterprise documents into a conversational chatbot. Put contracts, handbooks, policies and specs into a folder; employees or applications ask questions in plain language, and an AI agent reads your documents and answers with cited sources. > 把企业文档变成可以对话的聊天机器人：把合同、手册、制度、规范等文档放进一个文件夹，员工或应用系统就能用自然语言提问，AI 代理会查阅文档、给出带出处的回答。

  

---

What is this? / 这是什么？

An out-of-the-box enterprise knowledge Q&A robot. No index building, no vector database, no coding — 一个开箱即用的企业知识问答机器人。不需要建索引、不需要向量库、不需要写代码——

1. Add documents / 放文档：put enterprise docs (contracts, employee handbooks, reimbursement policies, IT policies, product specs...) into a folder 2. Ask / 提问：employees or internal systems ask in natural language, e.g. "How does annual leave work?" / 「年假怎么算？」 3. Answer / 回答：the AI agent reads your documents on the spot and answers with cited sources (which file, which line) / 现场查阅你的文档，给出准确回答并注明出处

Use cases: internal knowledge base Q&A, contract clause lookup, policy Q&A, system operation guides, onboarding assistance. 适用于：内部知识库问答、合同条款查询、制度答疑、系统操作说明、入职培训辅助等场景。

---

How to use (for ordinary users) / 普通用户怎么用

Option 1: ask directly in the DSH chat / 方式一：直接在 DSH 聊天界面问

Just ask in the DSH Web UI, e.g. / 在 DSH Web 界面直接提问即可，例如：

> "According to the enterprise documents, what is the annual leave policy?" / 「根据企业文档，员工的年假政策是什么？」

Option 2: via internal apps / chat UIs / web bots / 方式二：通过内部应用 / 聊天界面 / 网页机器人问

Any enterprise system, office software or web chat box can integrate. To the consumer it is just a "one question, one answer" conversation endpoint: 企业系统、办公软件、网页聊天框都可以接入。对使用方来说，就是一个「问一句、答一句」的对话接口：

# ask a question (wait for the answer) / 问一个问题（等待回答）
curl -X POST -H "Content-Type: application/json" \
  -d '{"prompt":"根据企业文档，酒店住宿报销上限是多少？"}' \
  "http://127.0.0.1:3080/workspace-api/task?wait=1"

Sample response / 返回示例：

{
  "ok": true,
  "task": {
    "status": "done",
    "result": "酒店报销上限：标准间每晚上限 HK$1200（出处：报销政策.md 第 3 行）",
    "exitCode": 0
  }
}

Real Q&A results (tested) / 典型问答效果（实测）

With 3 sample documents, answers returned within 10 seconds / 放 3 份示例文档，10 秒内返回：

---

Deployment (for admins) / 管理员怎么部署

Install / 安装

dsh plugin --profile web add dsh-workspace-api

Restart dsh web to activate / 重启 dsh web 后即生效。

Set the document folder / 指定企业文档目录

Defaults to the current DSH workspace; a dedicated folder is recommended / 默认使用 DSH 当前工作区；推荐专门指定一个文档目录：

WORKSPACE_API_ROOT=/srv/company-docs dsh web

> Put contracts, handbooks etc. into this folder (txt / md / PDF / Word / Excel supported); employees can then ask the bot. > 把合同、手册等文档放到这个目录（支持 txt / md / PDF / Word / Excel），员工就能向机器人提问了。 > Scanned PDFs must be OCR'd to text first / 扫描版 PDF 需先转成文字（OCR），AI 才能检索。

Expose it / 对外提供服务

TOKEN=your-secret-token dsh web

Callers must send the token / 调用方需带令牌：

curl -H "Authorization: Bearer your-secret-token" \
  "http://127.0.0.1:3080/workspace-api/task?wait=1" \
  -d '{"prompt":"..."}'

---

API Reference (for developers) / 面向开发者：API 参考

Served on the DSH GUI's own port (default 127.0.0.1:3080), prefix /workspace-api. 服务运行在 DSH 同端口（默认 127.0.0.1:3080），前缀 /workspace-api。 Every response uses {"ok": true, "data": ...} / {"ok": false, "error": ...}; CORS enabled. 所有响应统一为 {"ok": true, "data": ...} / {"ok": false, "error": ...}；已开启 CORS。

Endpoints / 常用端点

The question endpoint (core) / 提问接口（核心）

# async: returns a task id immediately / 异步提交：立即返回任务号
curl -X POST -H "Content-Type: application/json" \
  -d '{"prompt":"在 projects/Ams 里找导入 contract fee 的方法","timeoutMs":600000}' \
  "http://127.0.0.1:3080/workspace-api/task"
# → {"ok":true,"taskId":"...","status":"queued"}

# poll / 轮询结果
curl "http://127.0.0.1:3080/workspace-api/task/<taskId>"

# or synchronous (?wait=1) / 或同步等待
curl -X POST -H "Content-Type: application/json" \
  -d '{"prompt":"1+1=?","timeoutMs":120000}' \
  "http://127.0.0.1:3080/workspace-api/task?wait=1"

Request body / 请求体字段：

> Simple Q&A: ~3–5s. Code/document search: usually 1–3 min. Tasks run in a FIFO queue (single worker). > 简单问答约 3–5 秒；代码检索/文档问答通常 1–3 分钟。任务按队列顺序执行（单并发）。

Configuration / 配置项

---

Security / 安全说明

• Loopback-only by default; always set TOKEN when exposing beyond localhost / 默认仅监听 127.0.0.1；对外开放请务必配置 TOKEN
• Every path is realpath-checked against WORKSPACE_API_ROOT or registered workspaces; ../../etc-style traversal is rejected / 所有路径经真实路径校验，只能访问 WORKSPACE_API_ROOT 或已注册工作区，../../etc 之类一律拒绝
• Task agents have full DSH file capabilities; only expose to trusted callers / 任务代理拥有 DSH 完整文件能力，仅限可信调用方使用

---

How it works / 工作原理（简述）

employee / app --natural-language question--> /workspace-api/task
                                              | FIFO queue (single worker)
                                              v
                  dsh --profile headless "question" (fresh agent, cwd = docs folder)
                                              | on-the-spot search + reading
                                              v
                {"status":"done","result":"answer with cited sources"}

• Based on on-the-spot agent search; no index needed; best for tens to low-hundreds of documents / 基于 AI 代理现场检索，无需预建索引；文档少（几十份内）效果最佳
• Each question is a fresh session; no cross-question memory / 每次提问是全新会话，无跨问题记忆
• For very large corpora consider adding RAG / vector retrieval / 超大语料（上千份、文件名混乱）建议叠加 RAG 向量检索

---

Development & Publishing / 开发与发布

node --check lib/index.js                # syntax check / 语法检查
dsh plugin --profile web add link:$PWD   # local debug install / 本地调试安装
npm login && npm publish                 # publish to npm / 发布到 npm

After pushing to GitHub, add topics / 推送到 GitHub 后请添加 topic：dsh-plugin、deepseek-harness（社区市场会自动收录）。

---

License / 许可

MIT