Einstein: AI Co-scientist System

A multi-agent scientific hypothesis generation and evaluation system built on Microsoft's AutoGen framework.

Overview

Einstein is an implementation of an AI Co-scientist system, designed to help researchers generate, evaluate, and refine scientific hypotheses. The system uses a team of specialized AI agents, each focusing on different aspects of the scientific process:

• Supervisor: Orchestrates the research workflow and coordinates other agents
• Generation: Creates novel, testable scientific hypotheses
• Reflection: Reviews and critiques hypotheses with scientific rigor
• Ranking: Compares hypotheses through simulated scientific debates
• Evolution: Improves hypotheses based on feedback and combines strengths
• Meta-Review: Synthesizes insights and provides research overviews

The system follows a structured workflow:

1. Parse research goals into structured configurations
2. Generate initial hypotheses
3. Review and critique each hypothesis
4. Conduct a tournament to rank hypotheses (using an Elo-based rating system)
5. Improve the top-ranking hypotheses
6. Combine improved hypotheses into a unified hypothesis
7. Generate a comprehensive research overview

Installation

Install the package with pip:

pip install -e .

Or install the dependencies separately:

pip install -r requirements.txt

Usage

Quick Start

import asyncio
from einstein_pkg import AICoScientist

async def main():
    # Initialize the AI Co-scientist with your OpenAI API key
    co_scientist = AICoScientist(api_key="your-api-key")
    
    # Process a research goal
    research_goal = "Develop a novel hypothesis for the key factor contributing to ALS progression."
    results = await co_scientist.process_research_goal(research_goal)
    
    # Print the research overview
    print(results["research_overview"]["content"])
    
    # Add scientist feedback
    feedback = "I have a hypothesis that TDP-43 aggregation might be driven by impaired autophagy."
    contribution = await co_scientist.add_scientist_feedback(feedback)
    
    # Clean up resources
    await co_scientist.close()

if __name__ == "__main__":
    asyncio.run(main())

Command Line Example

Run the included example script:

python examples/basic_example.py --api-key your-api-key --model gpt-4-turbo

Or use a .env file with your API key and run without the --api-key parameter:

echo "OPENAI_API_KEY=your-api-key" > .env
python examples/basic_example.py

Package Structure

einstein_pkg/
├── einstein_pkg/           # Main package
│   ├── __init__.py         # Package initialization
│   ├── config.py           # Configuration handling
│   ├── main.py             # Main AICoScientist class
│   ├── memory.py           # Context memory persistence
│   ├── agents/             # Agent implementations
│   │   ├── __init__.py
│   │   ├── base.py         # Base agent class
│   │   ├── supervisor.py   # Research supervisor agent
│   │   ├── generation.py   # Hypothesis generation agent
│   │   ├── reflection.py   # Hypothesis review agent
│   │   ├── ranking.py      # Hypothesis ranking agent
│   │   ├── evolution.py    # Hypothesis improvement agent
│   │   └── meta_review.py  # Research synthesis agent
│   ├── models/             # Domain models
│   │   ├── __init__.py
│   │   └── research_config.py  # Research goal configuration
│   └── tools/              # AutoGen tools
│       ├── __init__.py
│       ├── literature_search.py  # Literature search tool
│       └── research_goal.py      # Research goal parsing tool
├── examples/               # Example scripts
│   └── basic_example.py    # Basic usage example
├── setup.py                # Package setup
└── requirements.txt        # Dependencies

Configuration

Einstein can be configured through:

1. Constructor parameters: Pass configuration directly when creating the AICoScientist instance
2. Environment variables: Set configuration in environment or .env file
3. Configuration dictionary: Use the configure() function to set global configuration

Common configuration options:

• openai_api_key: Your OpenAI API key
• default_model: Model to use (default: "gpt-4-turbo")
• context_memory_path: Path to store context memory (default: "context_memory.json")

Advanced Usage

Custom Memory Path

co_scientist = AICoScientist(context_memory_path="my_research_memory.json")

Clearing Memory

co_scientist.clear_memory()

Getting Latest Results

latest_results = co_scientist.get_latest_results()

License

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

Acknowledgments

This project was inspired by and implements concepts from the paper "Towards an AI Co-scientist for Accelerating Scientific Discoveries" and uses Microsoft's AutoGen framework for multi-agent communication.