How UltraRAG Turns RAG Deployment into a Zero‑Code, Multi‑Modal Powerhouse

Overview

UltraRAG is an open‑source Retrieval‑Augmented Generation (RAG) framework jointly developed by Tsinghua THUNLP and Northeastern University NEUIR. It integrates data construction, model fine‑tuning, and inference evaluation into a single, zero‑code workflow, targeting the three traditional pain points of RAG development: fragmented data pipelines, cumbersome retrieval tuning, and lack of unified evaluation.

Core capabilities

Zero‑code Web UI – A visual interface lets users create knowledge bases, slice and label domain data, launch fine‑tuning, and run evaluations without writing any code.

One‑click data construction and fine‑tuning – Built‑in methods such as KBAlign and RAG‑DDR automatically segment raw documents, generate pseudo‑labels, and prepare training data. The framework supports standard fine‑tuning (full‑parameter SFT) and parameter‑efficient LoRA adapters.

Multidimensional evaluation (RAGEval) – In addition to ROUGE‑L, RAGEval reports effective information recall and key‑knowledge coverage, providing a more robust assessment of answer quality.

Multimodal support (VisRAG) – VisRAG parses mixed‑media PDFs (tables, figures, formulas, CT‑scan screenshots) and incorporates extracted visual tokens into the knowledge base, enabling true multimodal retrieval.

Modular architecture – The system is organized into module, workflow, and function layers. Retrieval, encoder, and generator components can be swapped independently, facilitating both rapid product deployment and research prototyping.

Open‑source and containerizable – Licensed under Apache‑2.0, UltraRAG can be launched with a single docker‑compose command or via a Conda environment, without any commercial restrictions.

Evaluation example

On a legal‑domain benchmark, a vanilla RAG achieved 40.75 % ROUGE‑L. After applying the RAG‑DDR optimization, UltraRAG raised ROUGE‑L to 53.14 % and increased key‑clause recall by roughly 30 %.

Typical use cases

1. Enterprise digitalization (legal domain)

Deploy UltraRAG with docker‑compose up --build -d and open the Web UI at http://localhost:8843.

Import legal textbooks; the system automatically slices documents and generates labels using KBAlign.

Select the “RAG‑DDR” strategy and start retrieval plus fine‑tuning with a single click.

Run RAGEval – ROUGE‑L improves from 40.75 % to 53.14 % and key‑clause recall rises by 30 %.

Deployed lawyers receive instant, accurate statutory answers, achieving a 5× speedup over manual lookup.

2. Multimodal medical knowledge base

Upload a PDF containing text and CT‑scan screenshots via the knowledge‑base manager.

VisRAG extracts visual features (e.g., lesion locations, numeric indicators) and merges them with textual content into structured knowledge slices.

When building the retrieval chain, enable “multimodal retrieval”. Queries return both relevant text excerpts and matching imaging cases.

The generated answer combines textual and visual evidence, delivering more comprehensive diagnostic suggestions than text‑only RAG.

3. Research prototyping

Replace the default retrieval module with a custom algorithm using UltraRAG’s plug‑in interface.

Load built‑in evaluation datasets or import a custom set, then launch multidimensional evaluation with one click.

Compare effective information recall and key‑knowledge coverage between the new method and baseline models.

By eliminating repetitive data‑processing and evaluation code, researchers can focus on algorithmic innovation, roughly doubling experimental throughput.

Quick start – deployment options

Option 1: Docker (recommended)

# Clone the repository (optional)
git clone https://github.com/OpenBMB/UltraRAG.git
cd UltraRAG
# Build and start containers
docker-compose up --build -d

After the containers are running, open http://localhost:8843 to access the zero‑code Web UI.

Option 2: Conda environment

# Create a new Conda environment
conda create -n ultrarag python=3.10
conda activate ultrarag
# Install Python dependencies
pip install -r requirements.txt
# Download default models (resources/models)
python scripts/download_model.py
# Launch the Web UI
streamlit run ultrarag/webui/webui.py --server.fileWatcherType none

Visit http://localhost:8843 to begin using UltraRAG.

Repository

Project URL:

https://github.com/OpenBMB/UltraRAG