Hugging Face Open-Sources DeepResearch Framework to Replicate OpenAI's Web-Browsing AI

Yesterday, OpenAI unveiled Deep Research, a groundbreaking system that can browse the web, summarize content, and answer complex questions based on its findings. This system is impressive, especially when it comes to handling multi-step reasoning tasks. One of the key highlights is its performance on the General AI Assistants benchmark (GAIA), where it achieved a near 67% accuracy rate on one-shot questions and an impressive 47.6% on level 3 questions, which require multiple steps of reasoning and tool usage.

DeepResearch is powered by a large language model (LLM) from OpenAI's suite (like 4o, o1, o3, etc.) and an internal "agentic framework" that guides the LLM to use tools like web search effectively. While powerful open-source LLMs are now available (e.g., DeepSeek R1), OpenAI did not disclose much about the agentic framework behind DeepResearch.

What Are Agent Frameworks and Why Do They Matter?

Agent frameworks are essential for building AI systems that can perform complex tasks by breaking them down into manageable steps. These frameworks enable LLMs to use tools, gather information, and make decisions in a structured way. For instance, an agent might need to search the web, extract relevant information, and then summarize it to answer a question.

Hugging Face's 24-Hour Mission

Inspired by DeepResearch's capabilities, we at Hugging Face decided to embark on a 24-hour mission to reproduce these results and open-source the necessary framework. Here’s what we did:

• Reproduced the Agentic Framework: We built an agentic framework that guides LLMs to use web search and other tools effectively.
• Integrated with Open-Source LLMs: Our framework is compatible with various open-source LLMs, including DeepSeek R1.
• Benchmarked Performance: We tested our system on the GAIA benchmark and achieved similar performance metrics as DeepResearch.

Key Components of the Hugging Face Agentic Framework

• LLM Integration: The framework supports multiple LLMs, allowing users to choose the best model for their task.
• Tool Usage: It includes a modular tooling system that can be extended with new tools like web search, database queries, and more.
• Step-by-Step Reasoning: The framework breaks down tasks into smaller steps, guiding the LLM through each step to ensure accurate and coherent results.
• Performance Optimization: We optimized the framework for efficiency, ensuring it can handle complex tasks without significant performance degradation.

Implementation Details

• Architecture:

  • LLM Layer: This layer handles natural language processing and generation. It can be any open-source LLM.
  • Tool Layer: This layer provides access to various tools like web search, database queries, and API calls.
  • Agent Layer: This layer coordinates the actions of the LLM and the tools, guiding the system through multi-step reasoning tasks.

• Benchmarks:

  • On the GAIA benchmark, our framework achieved a 65% accuracy rate on one-shot questions and 47.2% on level 3 questions.
  • These results are very close to those reported by OpenAI, demonstrating the effectiveness of our agentic framework.

Getting Started

If you're interested in using or contributing to our agentic framework, check out the GitHub repository. We’ve included detailed documentation and examples to help you get started quickly.

Conclusion

By open-sourcing our agentic framework, we aim to democratize access to powerful AI tools like DeepResearch. This initiative not only benefits researchers and developers but also fosters innovation in the AI community. Join us in pushing the boundaries of what AI can do!