Assemble an AI expert panel to debate any question — with live web evidence.
圆桌智囊团 · 多 AI 角色协同讨论工作台
Drop in a topic. The AI assembles 4–8 expert personas with opposing stances, searches the web for live evidence, debates in multiple rounds, and delivers a synthesis. You listen, ask follow-ups, and decide.
| Feature | Description |
|---|---|
| Auto-assembled Panel | AI creates 4–8 distinct personas (background, stance, style) from your topic — no manual config |
| Live Web Evidence | Each speaker searches DuckDuckGo + Wikipedia before responding and cites sources inline |
| Private Knowledge Base | Upload PDF / Word / Markdown / TXT — AI retrieves relevant chunks during the debate |
| File Description Cards | One-sentence summary per document guides retrieval, reducing false hits |
| Evidence Trail | Every message shows its search queries, retrieved snippets, and source URLs |
| Any OpenAI-Compatible API | Works with OpenAI, DeepSeek, Moonshot, Anthropic, local Ollama models — just bring your API key |
| 100% Browser-Based | No server, no account, no data leaves your device (IndexedDB storage) |
| CN / EN UI | Full Chinese and English interface, auto-detects browser language |
→ Live Demo on Vercel
(Bring your own API key — enter it in Settings after opening)
The local launcher includes a proxy server for web search. Do not use a plain static server or evidence search will not work.
# Windows — double-click or run:
Start-Dev.bat
# Or manually:
powershell -ExecutionPolicy Bypass -File launcher/serve-static.ps1 -Root . -Port 4174Then open: http://127.0.0.1:4174/prototype-ui/
Your Topic
│
▼
① AI Moderator organizes the question and assigns expert roles
│
▼
② Each expert searches the web + your knowledge base for evidence
│
▼
③ Multi-round structured debate with citations
│
▼
④ Synthesis & conclusion
- Open Knowledge Base panel → upload documents (PDF / Word / Markdown / TXT)
- Write a File Description for each document (one sentence: what it covers, what questions it answers)
- Enable the Knowledge Base toggle before starting — relevant chunks are auto-injected into each speaker's context
Open Settings → Model Profiles and enter your API Key + Base URL.
Tested providers: OpenAI · Anthropic · DeepSeek · Moonshot (Kimi) · SiliconFlow · Ollama (local)
| Architecture | Vanilla JS single-file, zero framework, zero build step |
| Storage | IndexedDB (knowledge base, discussion history, evidence cache) |
| Web Proxy | api/proxy.js (Vercel Serverless) / launcher/serve-static.ps1 (local) |
| Retrieval | Chunked documents + TF-IDF keyword scoring + file description pre-filter |
| Deployment | Static files on Vercel; no backend required |
Early prototype — core features stable, actively iterating.
MIT © 2026 Roundtable Braintrust Contributors