Plugin: dsh-rag-kb — a local-first RAG knowledge base for DeepSeek Harness (Ollama, no cloud)

[AlowEnsoul]

Hi everyone! 👋

I built a local RAG knowledge base plugin for DeepSeek Harness (dsh). It adds a
draggable floating panel at the bottom-left of the dsh web UI where you can
drop your local documents — everything stays on your machine. No cloud, no
upload, no tracking.

[🎬 DEMO GIF: drag the panel, add files, ask "查一下知识库里关于 XX 的内容" — see it in action]

Why I built it

I wanted to give the dsh agent access to my own docs (notes, drafts, specs)
without sending anything to a third party. So the whole pipeline runs locally:

browser (ui-kb) ──fetch /kb──▶ host (rag-kb) ──curl──▶ Ollama /api/embed
    ▲                              │
    │                        chunking + cosine search
    │                              │
    └── JSON persistence ◀── dsh-kb-store.json

Features

• 🎈 Draggable floating entry — remembers its position, never blocks the UI
• 📄 Native Windows file picker — .txt/.md/.json/.py/.doc/.docx (Word text auto-extracted)
• 🔍 Semantic search — Ollama embeddings + cosine similarity, not keyword matching
• 🗂️ Multiple knowledge bases — independent indexes
• ⚙️ Configurable — embedding service URL, model, chunk size
• 🌗 Light / Dark / System themes
• 🖥️ Optional desktop console — a tiny EXE to start/stop harness & Ollama
• 🛠️ Agent-facing tools: kb_search / kb_index / kb_status + /kb HTTP route
• 📍 Chinese-path safe — paths are passed via base64 (this one's for the CN users 😄)

Quick start (full guide in README)

1. Copy packages/kb/rag-kb + packages/client/ui-kb into your dsh repo
2. Add 4 lines to tsconfig/bundle config (documented in README)
3. Merge config/cordis.patch.yml.example into your user patch
4. pnpm install && build, restart dsh web, ollama pull nomic-embed-text

What's next

• Demo video & screenshots (working on it)
• More file types & auto-sync folders
• Retrieval quality evaluation (comparing chunk strategies / embedding models)

This is my first open-source project — any feedback, issues, or ideas are
welcome! Repo: https://github.com/AlowEnsoul/dsh-rag-kb (MIT)

---

作者 AlowEnsoul · feedback: 2724749416@qq.com

中文版： 大家好 👋 我做了个 DeepSeek Harness 的本地 RAG 知识库插件：左下角可拖拽悬浮面板，本地文档切片→向量化→语义检索，全程本机运行（Ollama + 纯前端/主机端），无第三方云服务。支持多知识库、.doc/.docx 文字提取、三档主题、可选桌面控制台，并注册了 kb_search/kb_index/kb_status 三个 agent 工具。中文路径已用 base64 处理，对中文用户友好。这是我的第一个开源项目，欢迎一切反馈！仓库：https://github.com/AlowEnsoul/dsh-rag-kb（MIT）