UltraRAG

UltraRAG is an open-source RAG development framework built on the Model Context Protocol (MCP). It was launched jointly by THUNLP at Tsinghua University, NEUIR at Northeastern University, OpenBMB and AI9stars. It targets research exploration and industrial prototyping, and the latest documented release is UltraRAG 3.0.

What It Is

UltraRAG standardizes core RAG components such as the Retriever and Generation as independent MCP Servers. An MCP Client orchestrates them. Developers define a pipeline in a YAML file, including serial steps, loops and conditional branches, and run it with the ultrarag run command. New functionality is registered as function-level Tools so it can be reused across workflows.

UltraRAG UI

The project describes its UI as a visual RAG IDE for orchestration, debugging and demonstration. Its Pipeline Builder keeps a canvas view and a code view in sync, so pipeline parameters and prompts can be adjusted online. An AI assistant helps with pipeline design, parameter tuning and prompt generation. A pipeline can be converted into an interactive chat web UI with one command, and knowledge base management supports document Q&A.

Research and Evaluation Workflow

UltraRAG ships standardized evaluation workflows, mainstream research benchmarks and baseline integration with unified metric management. A case study interface lets users trace intermediate outputs of each pipeline step for error analysis. The documentation also covers calling UltraRAG components from Python code. A Deep Research demo pairs the pipeline with the AgentCPM-Report model to run multi-step retrieval and generate long survey reports.

Setup Path

UltraRAG installs from source with uv, using core, full or per-module extras, or runs from Docker images in CPU and GPU variants. The UI is served on port 5050 in the container. Running the sayhello example pipeline verifies the installation. The repository is licensed under Apache-2.0.