UltraRAG

UltraRAG is a Python-based MCP framework developed by Tsinghua University's THUNLP lab, Northeastern University, OpenBMB, and AI9Stars that lets developers build retrieval-augmented generation pipelines by composing independent MCP Servers for core components such as Retriever and Generator, then orchestrating them through an MCP Client using YAML configuration files. The framework supports sequential, loop, and conditional branch control structures, enabling iterative RAG logic with minimal code. It connects to LLMs including GPT, DeepSeek, Qwen, and vLLM-served models, and works with embedding models from Hugging Face sentence-transformers. A built-in Flask-based UI functions as a visual RAG IDE with a Pipeline Builder that synchronizes canvas construction and code editing in real time, plus a knowledge base management module for document question-answering. A standout feature is the AgentCPM-Report integration, an 8B on-device writing agent for localized deep research. Researchers prototyping novel RAG architectures and engineers building production retrieval systems both benefit from the low-code workflow and transparent reasoning visibility introduced in version 3.0.

What people ask about UltraRAG