# Welcome to Deep Research This is a research tool as a multi-agent AI research system designed to know what it knows (and doesnt know) when conducting research. This might seem like a small thing within research, but if you really think about it, this is the start of something much bigger. If the agents can understand what they dont know—just like a human—they can reason about what they need to learn. This has the potential to make the process of agents acquiring information much, much faster and in turn being much smarter. Click the image below to watch a demo of the tool and how the code works in detail: [](https://www.youtube.com/watch?v=mGET1RKXW3o) ## The Vision Currently, AI research agents have significant limitations - they often make assumptions, fabricate information, or miss crucial context. This project was heavily inspired by two groundbreaking works: 1. TICKing All the Boxes: Generated Checklists Improve LLM Evaluation and Generation by Jonathan Cook, it demonstrated how structured checklists can significantly improve LLM evaluation and generation quality. 2. NVIDIAs Nemotron-4 340B Reward Model, which pioneered techniques for reward modeling in the loop. Their work on reward-based learning has influenced our approach to quality assessment and continuous improvement in research tasks. ## System Architecture  This project is built on CAMEL-AIs Workforce system, which provides a robust framework for coordinating multiple AI agents. The workforce architecture enables our agents to work together seamlessly, with built-in task management and failure handling. Your feedback and contributions help make this project better! ## How It Works The system operates through 5 key components: ### 1. Recipes Think of these as detailed instructions for the research process. While currently manual, future versions will aim to automate recipe generation. Recipes specify: - Required information (ingredients) - Output format - Example outputs - Research parameters ### 2. Research Intelligence Planning Agent The first step in our research pipeline: - Analyzes input content against recipe requirements - Maps known vs unknown information - Creates structured research plans - Identifies which gaps can be filled through research ### 3. Deep Search Agent Our dedicated researcher: - Executes research plans - Uses Google search strategically - Verifies information from multiple sources - Documents findings and confidence levels ### 4. Report Creator Agent Our content synthesizer: - Combines original content with research findings - Follows recipe format requirements - Maintains clear sourcing - Highlights any remaining uncertainties ### 5. The Judge Agent Our final quality check: - Evaluates output completeness - Checks adherence to recipe - Validates information accuracy - Provides detailed quality metrics ## Prerequisites - Python 3.8+ - API Keys: - OpenAI API key (for GPT-4) - Anthropic API key (for Claude) - Google Custom Search API key - Google Custom Search Engine ID (cx) ## Installation 1. Clone the repository: git clone https://github.com/your-username/deep-research.git cd deep-research 2. Install required dependencies: pip install -r requirements.txt 3. Create a
