<2-3 sentences summarising what the project addresses and solves, the target audience, and the benefit.> This repository provides Pokee's 7B DeepResearch Agent, designed to answer complex questions by integrating real-time web search and content reading. It targets engineers, researchers, and power users seeking up-to-date, citation-rich research reports. The project offers a cost-effective API alternative to major providers, delivering comprehensive insights with no hidden fees or API key management.

How It Works

The PokeeResearch Agent leverages a 7B parameter model for scalable performance, integrating web search, content reading, and browsing tools. Its core approach involves multi-turn research, performing iterative web searches and content analysis to synthesize information. This design allows for comprehensive evaluation and high performance on complex reasoning tasks, utilizing the most current online data.

Quick Start & Requirements

• Hardware: Tested on a single 80GB A100 GPU; CUDA 12.8, NVIDIA driver 570.133.20. Multiple GPUs can accelerate inference.
• Software: Docker image provided (verlai/verl:app-verl0.5-transformers4.55.4-sglang0.4.10.post2-mcore0.13.0-te2.2). Python dependencies include colorlog, google-genai, vllm, httpx.
• API Keys: Requires Serper API, Jina API, Gemini API, and a HuggingFace Token.
• Setup: Pull Docker image, clone repository, install Python dependencies, and configure a .env file with API keys.
• Execution: Start the tool server (python start_tool_server.py), run benchmark evaluation (bash run.sh), or launch the CLI/GUI applications.
• Docs: API preview available at pokee.ai/deepresearch-preview. Model on Hugging Face. Paper available on arXiv.

Highlighted Details

• Performance: Achieves state-of-the-art results across 10 deep research benchmarks, including GAIA, BrowseComp, and TQ, outperforming other methods.
• Cost-Effectiveness: The associated API is advertised as up to 75% cheaper than OpenAI, Gemini, and Perplexity.
• Capabilities: Features multi-turn research, seamless tool integration for web search and content extraction, and research threads synthesis.
• Evaluation: Employs a robust evaluation framework using Gemini-2.5-Flash-lite on a dataset of 1,228 questions.

Maintenance & Community

Support is available via GitHub issues. Community engagement channels include Discord, LinkedIn, X (Twitter), and WeChat.

Licensing & Compatibility

The project is licensed under the Apache 2.0 license, which permits commercial use and integration with closed-source projects.

Limitations & Caveats

• Hardware Dependency: Requires a high-end GPU (80GB A100 tested), potentially limiting accessibility for users without specialized hardware.
• External API Reliance: Functionality is dependent on obtaining and configuring multiple third-party API keys (Serper, Jina, Gemini, HuggingFace).
• Setup Complexity: Docker-based deployment and the management of several API keys add complexity to the initial setup process.