README.md

metadata

title: Heavy Multi-Model 2.0
emoji: 🤖
colorFrom: blue
colorTo: purple
sdk: gradio
sdk_version: 4.44.1
app_file: app.py
pinned: false

🤖 Heavy Multi-Model 2.0 - AI Analysis System

Version 2.0 - Now with Web Search Integration!

A powerful multi-agent AI analysis system that leverages multiple AI models working in parallel to provide comprehensive, well-researched responses to complex queries.

🆕 What's New in v2.0

Web Search Integration with Tavily

• Real-time Information: Agents can now search the web for current data and facts
• Enhanced Analysis: Each agent performs web searches for their specialized questions
• Source Citation: Responses include references to web sources when using search
• Optional Feature: Enable/disable web search with a simple checkbox
• User-provided API Key: Bring your own Tavily API key for complete control

Features from v1.0

• Multi-Model Support: GPT-5, GPT-5.1, Gemini 3 Pro Preview, Gemini 2.5 Pro, Claude 4.5 Sonnet, Claude 4.5 Opus, GPT-4.1 Mini, Gemini 2.0 Flash, Llama 3.1 70B
• Flexible Configuration:
  • Single Model Mode: Use one model for all roles
  • Multi-Model Mode: Assign different models to each role
  • make-it-heavy Mode: Use the original implementation
• Parallel Execution: Multiple agents analyze different aspects simultaneously
• Web Interface: Beautiful Gradio-based UI with real-time updates

🏗️ Architecture

Multi-Agent Workflow

1. Orchestrator: Breaks down your query into specialized research questions
2. Agents (2-8): Each agent:
   • Receives a specialized question
   • (NEW) Optionally searches the web for relevant information
   • Analyzes their question from a unique perspective
   • Runs in parallel with other agents
3. Synthesizer: Combines all agent insights into a comprehensive response

Role-Based Model Assignment

Different AI models excel at different tasks:

• Orchestrator: Requires strong reasoning to generate meaningful questions
• Agents: Need analytical depth and diverse perspectives
• Synthesizer: Must excel at combining and organizing information

You can assign the best model to each role, or keep it simple with a single model for everything.

🚀 Getting Started

Prerequisites

• Python 3.8 or higher
• OpenRouter API key (get it from openrouter.ai/keys)
• (Optional) Tavily API key for web search (get it from tavily.com)

Installation

1. Navigate to this directory:

cd "Heavy 2.0"

2. Install dependencies:

pip install -r requirements.txt

3. Run the application:

python app.py

4. Open your browser to http://localhost:7860

📖 Usage

Basic Workflow

1. Enter API Keys:

   • OpenRouter API key (required)
   • Tavily API key (optional, for web search)

2. Enable Web Search (Optional):

   • Check the "Enable Web Search" box
   • Enter your Tavily API key
   • Agents will now search the web for each specialized question

3. Choose Configuration Mode:

   • Single Model: Use one model for all roles (simplest)
   • Multi-Model: Assign different models to each role (most flexible)
   • make-it-heavy: Use the original implementation (cost-effective)

4. Configure Analysis:

   • Number of agents (2-8): More agents = more perspectives
   • Show agent thoughts: Display individual agent analyses

5. Submit Your Query and watch the magic happen!

Example Queries

Perfect for web search:

• "What are the latest developments in quantum computing?"
• "What are the current best practices for React 18?"
• "What happened in the tech industry this week?"

Great for analysis:

• "What are the trade-offs between microservices and monolithic architecture?"
• "How should I design a scalable authentication system?"
• "What factors should I consider when choosing a database?"

🔧 Configuration

Model Options

All models are accessed through OpenRouter:

Model	Best For	Speed	Cost
GPT-5.1	Frontier reasoning + creativity	Medium	$$$$
GPT-5	Creative solutions	Fast	$$$
Gemini 3 Pro Preview	Multimodal depth & planning	Medium	$$$
Gemini 2.5 Pro	Synthesis & organization	Medium	$$
Claude 4.5 Opus	Frontier reasoning & software engineering	Medium	$$$$
Claude 4.5 Sonnet	Deep reasoning	Medium	$$$
GPT-4.1 Mini	Cost-effective analysis	Fast	$
Gemini 2.0 Flash	Quick responses	Very Fast	$
Llama 3.1 70B	Open source option	Medium	$

Web Search (Tavily)

When enabled, each agent:

1. Receives their specialized question
2. Searches the web using Tavily's advanced search
3. Gets an AI-generated summary + top 5 search results
4. Uses this context to enhance their analysis
5. Cites sources in their response

Search Depth: Advanced (configurable in tavily_search.py) Results per query: 5 (configurable)

📁 Project Structure

Heavy 2.0/
├── app.py                  # Entry point for Gradio app
├── config.yaml             # Configuration file
├── requirements.txt        # Python dependencies
├── src/
│   ├── __init__.py
│   ├── multi_web.py        # Gradio web interface
│   ├── multi_client.py     # Multi-model API client
│   ├── multi_orchestrator.py  # Workflow orchestrator
│   ├── multi_agent.py      # Individual agent logic
│   └── tavily_search.py    # NEW: Tavily web search integration
└── README.md

🔐 Security & Privacy

• API Keys: Never stored, only used during your session
• Web Search: Tavily searches are performed on-demand and not cached
• No Data Collection: Your queries and responses are not logged or saved

🛠️ Development

Extending the System

Add a new model:

# In src/multi_client.py
MODELS = {
    "your-model": {
        "provider": "openrouter",
        "model_id": "provider/model-name",
        "display_name": "Your Model Name"
    }
}

Customize web search:

# In src/tavily_search.py
def search(self, query: str, max_results: int = 5):
    # Adjust max_results, search_depth, etc.
    pass

Modify agent behavior:

# In src/multi_agent.py
def _build_system_prompt(self):
    # Customize agent instructions
    pass

📊 Performance Tips

1. Use web search selectively: Great for current events, not needed for general analysis
2. Adjust agent count: More agents = more perspectives but slower/costlier
3. Choose models wisely:
   • Quick answers: Gemini 2.0 Flash + GPT-4.1 Mini
   • Deep analysis: Claude 4.5 Sonnet + GPT-5
   • Cost-effective: GPT-4.1 Mini for all roles

🤝 Contributing

This is based on the original Heavy/make-it-heavy project. v2.0 adds:

• Tavily web search integration
• Enhanced UI with search controls
• Dynamic context enrichment for agents
• Source citation capabilities

📄 License

[Same as original project]

🙏 Acknowledgments

• Original Heavy project from Justin-Choo
• Tavily for web search API
• OpenRouter for multi-model access
• Gradio for the amazing UI framework

---

Built with ❤️ using Python, Gradio, and multiple AI models

For questions or issues, please refer to the original project or check the Tavily/OpenRouter documentation.