This project enhances the AutoGen framework to advance multi-agent conversational systems towards Artificial General Intelligence (AGI) capabilities. It targets researchers and developers looking to improve group chat dynamics, decision-making, and complex task proficiency in autonomous AI agents. The primary benefit is a more sophisticated and capable multi-agent system.

How It Works

The project modifies AutoGen classes to enable advanced group chat functionalities, including an "Agent Council" for decision-making and speaker selection based on a specific prompting technique. It supports conversation continuity by loading chat histories and fosters team awareness among agents, where each agent understands its role and those of its peers. Advanced RAG with RAG-fusion and LLM re-ranking via llama_index, along with built-in domain discovery for external knowledge, are key components.

Quick Start & Requirements

• Install dependencies: pip install -r requirements.txt
• Requires OpenAI API key and GitHub personal access token. Optional API keys for SerpApi and Google Programmable Search Engine.
• To run demo: python autogen_modified_group_chat.py
• See examples: python example_research.py, python example_rag.py
• Official blog post: project launch

Highlighted Details

• Enhanced group chat dynamics and agent council for decision-making.
• Integrated advanced RAG with RAG-fusion and LLM re-ranking.
• Built-in domain discovery for knowledge outside LLM training data.
• Supports conversation continuity and agent team awareness.

Maintenance & Community

• Project is actively seeking contributions.
• A blog post announcing the project launch is available.
• Community engagement is encouraged via GitHub stars.

Licensing & Compatibility

• MIT License.
• Permissive license allows commercial use, modification, distribution, and sublicensing without significant restrictions.

Limitations & Caveats

This project leverages agents with local code execution capabilities and relies on GPT-4 Turbo's extended context window, which can incur significant costs. Users should proceed with caution. Future work includes supporting more research resources, chat history overflow management, and potentially smaller context windows with open-source LLMs.