Trace2 Documentation

Welcome to the Trace2 documentation! Trace2 is a PyTorch-like library for training AI agents end-to-end using general feedback.

Quick Links

What is Trace2?

Trace2 is a new AutoDiff-like tool for training AI systems with:

• 🎯 General Feedback: Use rewards, natural language, test results, or any feedback
• 📊 Computation Graphs: Automatic tracing of execution like PyTorch
• 🔧 Real Python Code: Write actual functions, not strings
• ⚡ Multiple Optimizers: Choose between OptoPrime, OPRO, and TextGrad
• 🤖 LLM Agnostic: Works with OpenAI, Anthropic, and 100+ providers

Getting Started

pip install trace-opt

Then write your first optimizable function:

from opto.trace import bundle
from opto.optimizers import OptoPrime

@bundle(trainable=True)
def my_function(input):
    """This function will be optimized."""
    return input

optimizer = OptoPrime(my_function.parameters())

Documentation Structure

📚 Getting Started

Start here if you're new to Trace2. Learn installation, basic concepts, and configuration.

→ Get Started

🧠 Core Concepts

Deep dive into computation graphs, nodes, bundles, and feedback functions.

→ Learn Concepts

🚀 Optimizers

Compare and learn about OptoPrime, OPRO, and TextGrad optimizers.

→ Choose Optimizer

📖 Guides

Best practices, LLM backends, debugging, and production tips.

→ Read Guides

💡 Examples

Hands-on examples from code optimization to multi-agent systems.

→ See Examples

🎓 Tutorials

Step-by-step tutorials converted from Jupyter notebooks.

→ Start Tutorial

📘 API Reference

Complete Python API documentation auto-generated from source.

→ API Docs

Community

• GitHub - Source code and issues
• Discord - Community chat
• Paper - NeurIPS 2024
• Roadmap - Future plans

Citation

If you use Trace2 in your research, please cite:

@article{cheng2024trace,
  title={Trace is the Next AutoDiff: Generative Optimization with Rich Feedback, Execution Traces, and LLMs},
  author={Cheng, Ching-An and Nie, Allen and Swaminathan, Adith},
  journal={arXiv preprint arXiv:2406.16218},
  year={2024}
}