togethercomputer/open_deep_research

What it solves

Traditional web search falls short when you need deep, multi-hop reasoning on a complex topic. This project delivers comprehensive, well-cited research reports by automating the full research process — planning, searching, evaluating, and iterating — so you get a thorough, sourced document without hours of manual digging.

How it works

It runs an agentic LLM workflow with multiple self-reflection stages. The system plans an approach, searches the web via Tavily, evaluates the information it finds, and iterates until the report is complete. Every source is cited. You can run it from the command line with a config file, or use the included Gradio webapp. It requires API keys for Together, Tavily, and HuggingFace, and can generate output as PDF, HTML, a podcast version, and a visual table of contents image.

Who it’s for

Researchers, analysts, and anyone who needs long-form, cited reports on complex topics — especially those that demand multi-step reasoning. Developers can also extend the architecture because it’s designed to be flexible and community-contributed.

Highlights

• Multi-stage process with self-reflection for quality information gathering
• Source verification with citations for all information
• Extensible architecture for community extension
• Output formats: PDF, HTML, podcast, and visual table of contents image
• Runs via command line or Gradio webapp