UltraRAG

// Overview

UltraRAG is a RAG development framework whose architecture is built on the Model Context Protocol. Instead of a Python class hierarchy, each core RAG component — retriever, generation, corpus processing, evaluation — is packaged as an independent MCP server exposing function-level tools, and pipelines are assembled by an MCP client.

The practical consequence is that pipeline logic becomes configuration. Sequential steps, loops and conditional branches are expressed in YAML, so an iterative RAG workflow that would normally be a few hundred lines of orchestration code fits in a few dozen lines of config. Adding a new capability means registering a tool on a server rather than subclassing anything, which is what keeps components reusable across pipelines.

It is a joint project of THUNLP at Tsinghua University, NEUIR at Northeastern University, OpenBMB and AI9stars, and is aimed at both academic research and industrial prototyping. Alongside the pipeline engine it ships UltraRAG UI — a visual RAG IDE with a Pipeline Builder that keeps a canvas and the underlying code in two-way sync — plus standardized evaluation workflows over mainstream RAG benchmarks so experiments stay reproducible and comparable.

// What it does

• Low-code orchestration: sequential, loop and conditional control flow declared in YAML rather than written as Python glue
• Atomic MCP servers — retriever, generation, corpus and evaluation are independent services whose tools compose across pipelines
• Unified evaluation with built-in workflows and ready-to-run mainstream RAG benchmarks for reproducible comparisons
• UltraRAG UI: a visual pipeline builder with two-way sync between canvas and code, plus knowledge-base management and one-click conversion of a pipeline into a chat app
• Multimodal and vision-language support, with vLLM and Hugging Face Transformers backends
• Docker images for CPU and GPU when you would rather not build a local Python environment

// Getting started

The project recommends installing from source with uv, which manages the virtual environment and the locked dependency set. A Docker path is available if you prefer not to configure Python locally.

pip install uv==0.12.0
git clone https://github.com/OpenBMB/UltraRAG.git --depth 1
cd UltraRAG

# Core only (enough for UltraRAG UI)
uv sync

# Everything: retrieval, generation, corpus processing, evaluation
uv sync --all-extras

source .venv/bin/activate

ultrarag run examples/experiments/sayhello.yaml

docker pull hdxin2002/ultrarag:v0.3.0-base-cpu
docker run -it -p 5050:5050 hdxin2002/ultrarag:v0.3.0-base-cpu

Commands and code are distilled from the project's own documentation — always check the official repo for the latest.

// When to use it

• Run a standard RAG experiment on public benchmarks and compare it against baselines using the built-in evaluation workflow
• Prototype an iterative or branching RAG pipeline in YAML rather than writing bespoke orchestration code
• Reuse one retriever or reranker across several pipelines by exposing it once as an MCP server
• Turn a finished pipeline into an interactive chat demo with a knowledge base attached, without building a UI

// How UltraRAG compares

UltraRAG alongside other open-source rag frameworks & platforms tools AI/TLDR tracks, ranked by GitHub stars.