OpenBMB/UltraRAG

Less Code, Lower Barrier, Faster Deployment

     

简体中文  |  English

---

Latest News 🔥

• [2026.01.23] 🎉 UltraRAG 3.0 Released: Say no to "black box" development—make every line of reasoning logic clearly visible 👉 📖 Blog
• [2026.01.20] 🎉 AgentCPM-Report Model Released! DeepResearch is finally localized: 8B on-device writing agent AgentCPM-Report is open-sourced 👉 🤗 Model

Previous News

• [2025.11.11] 🎉 UltraRAG 2.1 Released: Enhanced knowledge ingestion & multimodal support, with a more complete unified evaluation system!
• [2025.09.23] New daily RAG paper digest, updated every day 👉 📖 Papers
• [2025.09.09] Released a Lightweight DeepResearch Pipeline local setup tutorial 👉 📺 bilibili · 📖 Blog
• [2025.09.01] Released a step-by-step UltraRAG installation and full RAG walkthrough video 👉 📺 bilibili · 📖 Blog
• [2025.08.28] 🎉 UltraRAG 2.0 Released! UltraRAG 2.0 is fully upgraded: build a high-performance RAG with just a few dozen lines of code, empowering researchers to focus on ideas and innovation! We have preserved the UltraRAG v2 code, which can be viewed at v2.
• [2025.01.23] UltraRAG Released! Enabling large models to better comprehend and utilize knowledge bases. The UltraRAG 1.0 code is still available at v1.

---

💡 About UltraRAG

UltraRAG is the first lightweight RAG development framework based on the Model Context Protocol (MCP) architecture design, jointly launched by THUNLP at Tsinghua University, NEUIR at Northeastern University, OpenBMB, and AI9stars.

Designed for research exploration and industrial prototyping, UltraRAG standardizes core RAG components (Retriever, Generation, etc.) as independent MCP Servers, combined with the powerful workflow orchestration capabilities of the MCP Client. Developers can achieve precise orchestration of complex control structures such as conditional branches and loops simply through YAML configuration.

🖥️ UltraRAG UI

UltraRAG UI transcends the boundaries of traditional chat interfaces, evolving into a visual RAG Integrated Development Environment (IDE) that combines orchestration, debugging, and demonstration.

The system features a powerful built-in Pipeline Builder that supports bidirectional real-time synchronization between "Canvas Construction" and "Code Editing," allowing for granular online adjustments of pipeline parameters and prompts. Furthermore, it introduces an Intelligent AI Assistant to empower the entire development lifecycle, from pipeline structural design to parameter tuning and prompt generation. Once constructed, logic flows can be converted into interactive dialogue systems with a single click. The system seamlessly integrates Knowledge Base Management components, enabling users to build custom knowledge bases for document Q&A. This truly realizes a one-stop closed loop, spanning from underlying logic construction and data governance to final application deployment.

https://github.com/user-attachments/assets/fcf437b7-8b79-42f2-bf4e-e3b7c2a896b9

✨ Key Highlights

🚀 Low-Code Orchestration of Complex Workflows Inference Orchestration: Natively supports control structures such as sequential, loop, and conditional branches. Developers only need to write YAML configuration files to implement complex iterative RAG logic in dozens of lines of code.	⚡ Modular Extension and Reproduction Atomic Servers: Based on the MCP architecture, functions are decoupled into independent Servers. New features only need to be registered as function-level Tools to seamlessly integrate into workflows, achieving extremely high reusability.
📊 Unified Evaluation and Benchmark Comparison Research Efficiency: Built-in standardized evaluation workflows, ready-to-use mainstream research benchmarks. Through unified metric management and baseline integration, significantly improves experiment reproducibility and comparison efficiency.	🎯 Rapid Interactive Prototype Generation One-Click Delivery: Say goodbye to tedious UI development. With just one command, Pipeline logic can be instantly converted into an interactive conversational Web UI, shortening the distance from algorithm to demonstration.

📦 Installation

We provide two installation methods: local source code installation (recommended using uv for package management) and Docker container deployment.

Method 1: Source Code Installation

We strongly recommend using uv to manage Python environments and dependencies, as it can greatly improve installation speed.

Prepare Environment

If you haven't installed uv yet, please execute:

## Direct installation
pip install uv==0.12.0
Download
curl -LsSf https://astral.sh/uv/0.12.0/install.sh | sh

Download Source Code

git clone https://github.com/OpenBMB/UltraRAG.git --depth 1
cd UltraRAG

Install Dependencies

Choose one of the following modes to install dependencies based on your use case:

A: Create a New Environment Use uv sync to automatically create a virtual environment and synchronize dependencies:

• Core dependencies: If you only need to run basic core functions, such as only using UltraRAG UI:

  uv sync
  

• Full installation: If you want to fully experience UltraRAG's retrieval, generation, corpus processing, and evaluation functions, please run:

  uv sync --all-extras
  

This is the recommended team setup: every MCP server shares the same .venv and uv.lock. Linux GPU dependencies are locked to CUDA 12.9, including the official vLLM cu129 wheel.

• On-demand installation: If you only need to run specific modules, keep the corresponding --extra as needed, for example:

  uv sync --extra retriever   # Retrieval module only
  uv sync --extra generation  # Generation module only
  

Once installed, activate the virtual environment:

# Windows CMD
.venv\Scripts\activate.bat
Windows Powershell
.venv\Scripts\Activate.ps1
macOS / Linux
source .venv/bin/activate

B: Install into an Existing Environment To install UltraRAG into your currently active Python environment, use uv pip:

# Core dependencies
uv pip install -e .
Full installation
uv pip install -e ".[all]"
On-demand installation
uv pip install -e ".[retriever]"

Method 2: Docker Container Deployment

If you prefer not to configure a local Python environment, you can deploy using Docker.

Get Code and Images

# 1. Clone the repository
git clone https://github.com/OpenBMB/UltraRAG.git --depth 1
cd UltraRAG
2. Prepare the image (choose one)
Option A: Pull from Docker Hub
docker pull hdxin2002/ultrarag:v0.3.0-base-cpu # Base version (CPU)
docker pull hdxin2002/ultrarag:v0.3.0-base-gpu # Base version (GPU)
docker pull hdxin2002/ultrarag:v0.3.0          # Full version (GPU)
Option B: Build locally
docker build -t ultrarag:v0.3.0 .

Start the Container

# Start the container (Port 5050 is mapped by default)
docker run -it --gpus all -p 5050:5050 <docker_image_name>

Note: After the container starts, UltraRAG UI will run automatically. You can directly access http://localhost:5050 in your browser to use it.

Verify Installation

After installation, run the following example command to check if the environment is normal:

ultrarag run examples/experiments/sayhello.yaml

If you see the following output, the installation is successful:

Hello, UltraRAG v3!

🚀 Quick Start

We provide complete tutorial examples from beginner to advanced. Whether you are conducting academic research or building industrial applications, you can find guidance here. Welcome to visit the Documentation for more details.

🔬 Research Experiments

Designed for researchers, providing data, experimental workflows, and visualization analysis tools.

• Getting Started: Learn how to quickly run standard RAG experimental workflows based on UltraRAG.
• Evaluation Data: Download the most commonly used public evaluation datasets in the RAG field and large-scale retrieval corpora, directly for research benchmark testing.
• Case Analysis: Provides a visual Case Study interface to deeply track each intermediate output of the workflow, assisting in analysis and error attribution.
• Structured Debugging Guide (Chinese): When answers look suspicious, retrieval hits are unstable, the reasoning chain drifts, or post-deployment behavior is abnormal, troubleshoot across four layers — input & retrieval, reasoning & planning, state & context, and deployment & runtime.
• Code Integration: Learn how to directly call UltraRAG components in Python code to achieve more flexible customized development.

🛠️ Demo Systems

Designed for developers and end users, providing complete UI interaction and complex application cases.

• Quick Start: Learn how to start UltraRAG UI and familiarize yourself with various advanced configurations in administrator mode.
• Deployment Guide: Detailed production environment deployment tutorials, covering the setup of Retriever, Generation models (LLM), and Milvus vector database.
• Deep Research: Flagship case, deploy a Deep Research Pipeline. Combined with the AgentCPM-Report model, it can automatically perform multi-step retrieval and integration to generate tens of thousands of words of survey reports.

🤝 Contributing

Thanks to the following contributors for their code submissions and testing. We also welcome new members to join us in collectively building a comprehensive RAG ecosystem!

You can contribute by following the standard process: Fork this repository → Submit Issues → Create Pull Requests (PRs).

⭐ Support Us

If you find this repository helpful for your research, please consider giving us a ⭐ to show your support.

💬 Contact Us

• For technical issues and feature requests, please use GitHub Issues.
• For questions about usage, feedback, or any discussions related to RAG technologies, you are welcome to join our WeChat group, Feishu group, and Discord to exchange ideas with us.
• If you have any questions, feedback, or would like to get in touch, please feel free to reach out to us via email at yanyk.thu@gmail.com

WeChat Group

Feishu Group

Discord

📖 Publications

Papers

1. Shi Yu, Chaoyue Tang, Bokai Xu, Junbo Cui, Junhao Ran, Yukun Yan, Zhenghao Liu, Shuo Wang, Xu Han, Zhiyuan Liu, Maosong Sun. (2025) VisRAG: Vision-based Retrieval-augmented Generation on Multi-modality Documents. arXiv:2410.10594 and In Proceedings of the Thirteenth International Conference on Learning Representations (ICLR 2025).

2. Xinze Li, Sen Mei, Zhenghao Liu, Yukun Yan, Shuo Wang, Shi Yu, Zheni Zeng, Hao Chen, Ge Yu, Zhiyuan Liu, Maosong Sun, Chenyan Xiong. (2025) RAG-DDR: Optimizing Retrieval-Augmented Generation Using Differentiable Data Rewards. arXiv:2410.13509 and In Proceedings of the Thirteenth International Conference on Learning Representations (ICLR 2025).

3. Kunlun Zhu, Yifan Luo, Dingling Xu, Yukun Yan, Zhenghao Liu, Shi Yu, Ruobing Wang, Shuo Wang, Yishan Li, Nan Zhang, Xu Han, Zhiyuan Liu, Maosong Sun. (2025) RAGEval: Scenario Specific RAG Evaluation Dataset Generation Framework. arXiv:2408.01262 and In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025).

4. Ruobing Wang, Qingfei Zhao, Yukun Yan, Daren Zha, Yuxuan Chen, Shi Yu, Zhenghao Liu, Yixuan Wang, Shuo Wang, Xu Han, Zhiyuan Liu, Maosong Sun. (2025) DeepNote: Note-Centric Deep Retrieval-Augmented Generation. arXiv:2410.08821 and In Findings of the Association for Computational Linguistics: EMNLP 2025.

Models

1. Yishan Li, Wentong Chen, Yukun Yan, Mingwei Li, Sen Mei, Xiaorong Wang, Kunpeng Liu, Xin Cong, Shuo Wang, Zhong Zhang, Yaxi Lu, Zhenghao Liu, Yankai Lin, Zhiyuan Liu, Maosong Sun. (2026) AgentCPM-Report: Interleaving Drafting and Deepening for Open-Ended Deep Research. arXiv:2602.06540.

2. OpenBMB. MiniCPM-Embedding-Light. Hugging Face Model Card.