Agent Debate（多智能体辩论赛）

Multi-Agent Debate Demo (FastAPI + LangGraph + LangChain OpenAI)

A hands-on demo for multi-agent debate-style reasoning, easy to run and easy to extend.

---

✨ Features

🧠 AI vs AI Multi-Agent Debate

• A judge plus 4 debaters for the Pro side and 4 for the Con side (9 agents in total).
• Each position (1st–4th speaker) has a different role and persona, fully configurable.
• Debate flow is orchestrated via a directed graph / FSM built with LangGraph.
• Supports NDJSON streaming for real-time UI updates.

---

👤 Human vs AI Debate

• You can freely choose to argue for the Pro or Con side.
• The AI side cycles through 4 debaters: opening, rebuttal, examples, and summary.
• Chat-style interaction with multi-turn debate rounds.

---

🖥 Built-in Frontend

• Comes with a minimal HTML page (index.html), no separate frontend project needed.
• Uses fetch + NDJSON streaming to update debate turns progressively.
• Default static assets (avatars, etc.) live under app/static/.

---

🧩 Tech Stack

• Backend
  • FastAPI
  • LangGraph
  • LangChain / LangChain OpenAI
• Model Providers
  • OpenAI / DeepSeek / DashScope (Qwen) / Kimi / BigModel (GLM) and others
• Frontend
  • HTML + JavaScript

---

📁 Project Layout

.
├─ requirements.txt
├─ README.md
├─ .env                           # Stores model API keys & base URLs
└─ app/
   ├─ __init__.py
   ├─ api.py                      # FastAPI entrypoint
   ├─ agent.py                    # DebateState / AgentRole / speak_with_role
   ├─ graph.py                    # LangGraph FSM for debate flow
   ├─ config.py                   # Model profiles & personas
   ├─ demo.py                     # CLI demo
   ├─ index.html                  # Frontend page
   └─ static/                     # Static assets (avatars, etc.)

---

🧱 Environment

• Python 3.10+ is recommended (ideally inside a virtual environment).
• Requires network access to the model providers you want to use (OpenAI / DeepSeek / DashScope / Kimi / BigModel, etc.).

---

📦 Installing Dependencies

1️⃣ Create a virtualenv

python -m venv .venv
# Windows
# .venv\Scripts\activate
# macOS / Linux
# source .venv/bin/activate

2️⃣ Install requirements

pip install -r requirements.txt

---

🔧 Configure Environment Variables

All model settings live in app/config.py under MODEL_PROFILES:

MODEL_PROFILES = {
    "gpt4.1": {
        "model": "gpt-4.1",
        "api_key_env": "OPENAI_API_KEY",
        "base_url_env": "OPENAI_BASE_URL",
        "label": "OpenAI GPT-4.1",
        "group": "OpenAI",
    },
    "deepseek-chat": {
        "model": "deepseek-chat",
        "api_key_env": "DEEPSEEK_API_KEY",
        "base_url_env": "DEEPSEEK_BASE_URL",
        "label": "DeepSeek Chat",
        "group": "DeepSeek",
    },
    "deepseek-reasoner": {
        "model": "deepseek-reasoner",
        "api_key_env": "DEEPSEEK_API_KEY",
        "base_url_env": "DEEPSEEK_BASE_URL",
        "label": "DeepSeek Reasoner",
        "group": "DeepSeek",
    },
    "qwen3-max": {
        "model": "qwen3-max",
        "api_key_env": "DASHSCOPE_API_KEY",
        "base_url_env": "DASHSCOPE_BASE_URL",
        "label": "Qwen3-Max",
        "group": "DashScope",
    },
    "kimi-k2-turbo-preview": {
        "model": "kimi-k2-turbo-preview",
        "api_key_env": "MOONSHOT_API_KEY",
        "base_url_env": "MOONSHOT_BASE_URL",
        "label": "Kimi K2 Turbo Preview",
        "group": "Kimi",
    },
    "glm-4.5": {
        "model": "glm-4.5",
        "api_key_env": "BIGMODEL_API_KEY",
        "base_url_env": "BIGMODEL_BASE_URL",
        "label": "GLM-4.5",
        "group": "BigModel",
    },
}

You can provide all keys & base URLs via environment variables:

• OPENAI_API_KEY
• OPENAI_BASE_URL
• DEEPSEEK_API_KEY
• DEEPSEEK_BASE_URL
• DASHSCOPE_API_KEY
• DASHSCOPE_BASE_URL
• MOONSHOT_API_KEY
• MOONSHOT_BASE_URL
• BIGMODEL_API_KEY
• BIGMODEL_BASE_URL

---

📄 Use a .env File (Recommended)

Create a .env file at the project root, for example:

# OpenAI
OPENAI_API_KEY=sk-xxxx
OPENAI_BASE_URL=https://api.openai.com/v1

# DeepSeek
DEEPSEEK_API_KEY=ds-xxxx
DEEPSEEK_BASE_URL=https://api.deepseek.com

# DashScope (Qwen)
DASHSCOPE_API_KEY=ds-xxxx
DASHSCOPE_BASE_URL=https://dashscope.aliyuncs.com/compatible-mode/v1

# Kimi Moonshot
MOONSHOT_API_KEY=ms-xxxx
MOONSHOT_BASE_URL=https://api.moonshot.cn/v1

# BigModel (GLM)
BIGMODEL_API_KEY=glm-xxxx
BIGMODEL_BASE_URL=https://open.bigmodel.cn/api/paas/v4

config.py already calls load_dotenv(), so values from .env are loaded automatically.

---

🖥 Command-Line Demo

Try an AI vs AI debate quickly in your terminal:

python -m app.demo

You will be prompted to choose, for example:

• Debate topic
• Rebuttal rounds
• Model profiles and other options

---

🚀 Run the Backend Server

Make sure your working directory is the project root, then run:

uvicorn app.api:app --reload

• Default host: http://127.0.0.1:8000
• You can change host/port with --host / --port.

---

💡 Use the Frontend Page

Once uvicorn is running, open:

http://127.0.0.1:8000/

On this simple page, you can:

• Choose the debate mode (AI vs AI / Human vs AI)
• Choose model profiles and personas
• Watch the debate stream in real time

---

🧩 Tips for Customization

• In app/config.py you can:
  • Add or edit MODEL_PROFILES to connect your own model endpoints.
  • Customize personas, tone, and responsibilities of each debater.
• In app/graph.py you can:
  • Change debate rounds or add new phases (e.g., free debate).
• In index.html you can:
  • Replace the minimal UI with any framework you prefer (Vue / React / Svelte / Tailwind, etc.).

---

✅ License

This project is licensed under the MIT License.