Turn your documents into a cited, ask-anything knowledge base. An open-source RAG platform — and the reference application for the OpenGenie AI Stack.
-1a7f37){kind=icon}
Quick Start · Agent install guide · Runbook · OpenGenie Stack
English ｜ 繁體中文
Upload PDFs → split → convert (Docling) → chunk + metadata/keywords → ingest to Qdrant → chat through Open WebUI, with downloadable source citations in every answer. Web UI, fully bilingual (EN / 中文), no code required.
This app is the application layer; the OpenGenie AI Stack is the infrastructure underneath (n8n · Qdrant · Open WebUI · Docling · PostgreSQL · Redis · FileBrowser). Deploy OpenGenie first, then plug this in — see the Connect to OpenGenie panel in the overview above. TigerAI RAG Community shows how the stack becomes a real, working product.
- Build knowledge bases from PDFs: server-side split → Docling Markdown → JSON chunking → metadata + L1/L2/L3 keyword dictionary → Qdrant ingest
- Rule Builder — generate a domain system-prompt from sample documents (A/B/C → pick or synthesize)
- Chat verification & QC — coverage check (source ↔ vector DB) and dual-AI answer audit
- Cloud chat apps — wire an Open WebUI app to an n8n query webhook in two clicks
- Citations with download links — every answer lists the source files used, each downloadable
- Bilingual UI — toggle EN / 中文 anytime
Hand it to your AI coding agent — no manual setup.
Open Claude Code / Antigravity / Codex and give it two URLs:
- https://github.com/TigerAI-Taiwan/OpenGenie-AI-Stack — the infrastructure
- this repo — the application
The agent reads both and installs + wires them end-to-end. The only thing it asks you for is a Cloud AI API key (any OpenAI-compatible endpoint).
📖 The agent follows AGENTS.md (entry) and INSTALL.md (the step-by-step runbook). Images are published multi-arch (amd64 + arm64, incl. NVIDIA Grace/GB10) at
ghcr.io/tigerai-taiwan/tigerai-rag-{nginx,backend,splitter}.
Prefer to drive it yourself? Follow INSTALL.md — discover the OpenGenie stack → docker compose up → import n8n workflows → install the Open WebUI pipes → set your Cloud AI key.
Env vars worth knowing (full list in
.env.example, see also INSTALL.md §2):
TIGERAI_EDITION=community— edition tag (do not change; Pro/Enterprise images use their own value).N8N_HOST=n8n-main— canonical OpenGenie main n8n container name (override if your stack uses a different name).SETTINGS_SECRET/TRANSIT_SECRET— MUST be set to strong random strings in production (openssl rand -hex 32); leaving them blank works for local dev only — backend will auto-generate per-restart values that break multi-instance and persist nothing.PG_PASSWORD/FB_PASSWORD— defaults are for local dev only; change in production (FileBrowser enforces ≥12 chars).
如果你是被指派來安裝這個專案的 AI agent(Claude Code / Antigravity / Gemini / Codex…),按以下順序讀這些文件,不要只看 README 就動手:
| # | 文件 | 為什麼要讀 | 一句話重點 |
|---|---|---|---|
| 1 | AGENTS.md | 安裝入口、兩階段 check-first 流程 | 先 --check-only dry-run 再實裝;只跟人要「無法用 API 拿到的密鑰」 |
| 2 | STACK-REFERENCE.md §4 | 強制 probe SOP | docker ps + docker network inspect 拿真實值 |
| 3 | INSTALL.md | 逐步 runbook | RAG/ 目錄自動處理、allowed_formats、port 衝突等細節 |
| 4 | LESSONS_LEARNED.md | 踩過的坑 + 修法 | fresh install 必中的 bug 全記在這(v1.0.10-25),含 n8n API key 必須含 credentials:read、Cloud AI 別用 reasoning model 等 |
| 5 | .env.example | 設定欄位參考 | 每個 env 的用途 + 哪些是必填 / 哪些有安全警告 |
最容易踩的 3 個雷(全寫在上面文件裡,這裡先警告):
- 不要套 STACK-REFERENCE 的預設容器名 — 你的機器可能是
open-webui(非openwebui-main)、docling-serve(非docling)、某些 service 不在同一個 network。必須 probe。 - n8n API key 必須含
credentials:read權限 — 否則 OpenAI 憑證綁不上、workflow 無法 active(deploy 會 fail-stop 提醒你)。 - Cloud AI model 用
gpt-4o-mini類標準模型 — 別用o1/gpt-5.xreasoning model(參數限制如不接temperature/max_tokens會讓訓練 chain 失敗)。
| Community (this repo) | Business | |
|---|---|---|
| AI | Cloud (any OpenAI-compatible endpoint) | + On-premise / local (Ollama, llama.cpp, vLLM) |
| Retrieval | Vector + metadata filtering | + Reranking, multimodal catalog RAG, trustworthy QC |
| Cross-document | Per-document Q&A (vector) | + Knowledge graph (cross-report, multi-hop relations) |
| Training | Direct training | + Scheduling queue |
| Price | Free, Apache-2.0 | Commercial |
Need local / on-prem AI, reranking, multimodal catalog RAG, or a cross-document knowledge graph? Those are TigerAI Advanced RAG Business features — get in touch.
Apache-2.0 — code is yours to use, fork, modify and redistribute freely.
TigerAI, the TigerAI logo, and product names such as TigerAI Advanced RAG, TigerAI Advanced RAG Community, and TigerAI Advanced RAG Business are trademarks of TigerAI Taiwan. The Apache-2.0 license covers the source code only — it does not grant rights to the name or logo.
You can fork and build on this codebase freely; please rebrand if you redistribute a modified version commercially or host it as a paid service. See TRADEMARK.md for the full policy (it's short and friendly — same model as Kubernetes / Docker / Grafana / PostgreSQL).
開源、網頁式的 RAG 知識庫建置平台,同時是 OpenGenie AI Stack 的官方範例應用 —— 示範如何把這套自架 AI 基礎設施變成一個真正可用的產品。
上傳 PDF → 切檔 → Docling 轉檔 → 切塊 + metadata/關鍵字 → 灌入 Qdrant → 透過 Open WebUI 問答,每個答案都附可下載的原始檔出處。全程網頁操作、中英雙語、免寫程式。
與 OpenGenie 的關係:OpenGenie 提供引擎(n8n / Qdrant / Open WebUI / Docling / PostgreSQL / Redis / FileBrowser);本專案是跑在上面的應用層,完整示範這套 stack 怎麼被運用。
安裝(免動手):打開 AI coding agent(Claude Code / Antigravity / Codex),把兩個 URL 丟給它 —— ① OpenGenie AI Stack(基礎設施)② 本 repo(應用)。Agent 會自己讀完、安裝並串接好,人不用介入;只會跟你要一把 Cloud AI API key。詳見 AGENTS.md / INSTALL.md。映像檔為 multi-arch(amd64 + arm64,含 NVIDIA Grace/GB10),位於 ghcr.io/tigerai-taiwan/tigerai-rag-*。
版本差異:Community(本 repo,免費、Apache-2.0、雲端 AI)= 陽春雲端版;地端/on-prem、Rerank 重排、多模態型錄 RAG、可信 QC、排程佇列、跨文件知識圖譜(跨報告多跳交叉查詢) 屬付費版 —— 需要請洽詢。
商標(Trademark):TigerAI、TigerAI logo 與「TigerAI Advanced RAG」等產品名稱為 TigerAI Taiwan 的商標。Apache-2.0 只授權程式碼,不授權品牌與 logo。可以自由 fork / 修改 / 使用程式碼;但若你修改後對外散佈或商業化(如 SaaS 服務),請改用自己的名稱與 logo(如「MyCompany RAG (fork of TigerAI)」這樣標示來源 OK)。完整政策見 TRADEMARK.md(與 Kubernetes / Docker / Grafana / PostgreSQL 同模式)。