Recently, the THUNLP Lab at Tsinghua University, the NEUIR Lab at Northeastern University, OpenBMB, and AI9Stars jointly released UltraRAG 2.0, which is the first retrieval-augmented generation (RAG) framework designed based on the Model Context Protocol (MCP) architecture.
This framework aims to simplify the construction of RAG systems, allowing researchers to quickly implement complex multi-stage reasoning systems within a short period. The highlight of UltraRAG 2.0 is that users can easily declare complex logic such as serial, loop, and conditional branches by writing YAML files, significantly reducing the amount of code and lowering the implementation barrier.
In the current trend of RAG development, many systems have gradually integrated advanced features such as adaptive knowledge organization, multi-turn reasoning, and dynamic retrieval, represented by projects like DeepResearch and Search-o1. However, these complex features also bring high engineering costs for developers, limiting the rapid iteration and replication of new ideas. UltraRAG 2.0 was born to address this issue, achieving flexible function calls and extensions by encapsulating core components of RAG into independent MCP servers.
Specifically, compared to previous implementations, UltraRAG 2.0 has significantly reduced the amount of code. For example, the official implementation of the classic method IRCoT requires nearly 900 lines of code, while UltraRAG 2.0 can achieve the same functionality with about 50 lines of code. Half of the code is used for workflow orchestration in YAML pseudo-code, greatly reducing the development barrier. The framework supports building multi-stage reasoning processes through a concise declarative approach, making complex reasoning logic no longer require lengthy manual coding.
UltraRAG 2.0 also supports advanced features such as dynamic retrieval, conditional judgment, and multi-turn interaction, enabling researchers to quickly build high-performance experimental platforms to meet the needs of complex multi-hop questions. Its performance is approximately 12% higher than traditional Vanilla RAG. The design of this system aims to save researchers time and effort in engineering implementation, allowing them to focus more on algorithm innovation and experimental design.
Additionally, the MCP architecture of UltraRAG 2.0 allows seamless reuse between different modules and supports flexible expansion and integration of modules, making it convenient for researchers to quickly adapt to new models and algorithms. This design greatly improves the efficiency and reproducibility of system development.
GitHub: https://github.com/OpenBMB/UltraRAG
Project Homepage: https://openbmb.github.io/UltraRAG
Key Points:
🌟 UltraRAG 2.0 was jointly developed by Tsinghua University and Northeastern University, aiming to simplify the process of building complex RAG systems.
🛠️ Users can implement complex reasoning logic by writing YAML files, significantly reducing the amount of code and the development barrier.
📈 Compared to traditional methods, UltraRAG 2.0 improves performance by about 12%, suitable for advanced features such as multi-turn reasoning and dynamic retrieval.
