OpenSeeker addresses the challenge of democratizing advanced search agent capabilities by fully open-sourcing its training data and models. It targets researchers and developers, enabling them to build, evaluate, and deploy sophisticated search agents for complex information-seeking tasks. The project's key benefit is providing an accessible, state-of-the-art system that was developed by a purely academic team.

How It Works

The system leverages a fine-tuned Qwen3-30B-A3B-Thinking-2507 large language model, trained on 11.7K examples. Its core approach involves integrating LLM-driven decision-making with tool usage, specifically search and web visit functionalities, to navigate and extract information from the web. This methodology allows for complex query decomposition and execution, aiming for state-of-the-art performance on frontier search benchmarks.

Quick Start & Requirements

Installation involves cloning the repository and setting up a Python 3.10 Conda environment (conda create --name openseeker python=3.10, conda activate openseeker, pip install -r requirements.txt). Model deployment requires git-xet for downloading the OpenSeeker-v1-30B-SFT model from Hugging Face, followed by updating the MODEL_PATH in run_openseeker.sh. API endpoints and keys are configured via setup_env.sh. Usage includes generating answers with python eval/generate_answer.py and evaluating results with `python eval/eval