God LLM 🤖

A framework for recursive thought expansion and exploration using Large Language Models.

Inspired by the article I wrote Let's Build a God LLM which explores the concept of recursive thought expansion and deep exploration using language models.

🌟 Features

Core Functionality

• Recursive Thought Expansion: Generate and explore interconnected thoughts while maintaining context
• Context Preservation: Maintains parent context during deep exploration
• Relevance Filtering: Automatically removes irrelevant thought branches
• Detailed Reporting: Generate comprehensive analysis of thought patterns
• Debug Mode: Conditional logging for development (DEBUG=True|False)
• Enhanced Visualization: Better representation of thought trees

Advanced Capabilities

• Grounded Prompting: Improved prompt templates to maintain focus
• Context Management: Efficient handling of conversation history
• Custom Visualization: Methods for displaying thought hierarchies

🛠️ Upcoming Features

High Priority

• Tool Integration
  • Implement with_tools() method for LLM classes
  • Support for Tavily integration
  • RAG (Retrieval Augmented Generation) implementation
  • Function calling in BaseLLM inheritor classes
  • Off-the-shelf tool library
  • Custom tool creation framework

Technical Improvements

• Fix infinite expansion in expand() method
• Implement comprehensive test suite
• Enhanced graph visualization
• Better divergence control in thought expansion

🚀 Getting Started

from god_llm.core.god import God
from god_llm.plugins.groq import ChatGroq

llm = ChatGroq(
    model_name="llama-3.1-70b-versatile",
    api_key="hello", #Dummy API key
)
god = God(llm=llm)

god.expand("Why?")

📘 Usage Examples

Basic Thought Expansion

god.report()

Using Tools

# Coming soon
llm = llm.with_tools([
    TavilyTool(),
    CustomTool(),
    RAGTool()
])

🤝 Contributing

Contributions are welcome! Please check out our contribution guidelines before getting started.

🐛 Known Issues

• expand() method may run indefinitely in certain cases
• Graph visualization can diverge from initial prompt
• Deep expansion may lose context of original prompt

📝 License

This project is licensed under the MIT License - see the LICENSE file for details.

🙏 Acknowledgments

• Original concept inspired by Islam Hachimi's article
• All contributors and feedback providers

Built with ❤️ for better thought exploration

When finding path, also account for relations, combine hierarchical (parent-child) and lateral (relations) nodes for retrieval, give higher weight for parent-child relationships

Allow to ask the right questions, that is the true wisdom of the God LLM

Implement function calling to only retreive questions

FIx errors in scoring