How UltraRAG Turns Complex RAG Deployment into a One‑Click, No‑Code Process

Overview

UltraRAG is an open‑source framework for Retrieval‑Augmented Generation (RAG) that integrates knowledge‑base construction, retrieval optimization, model fine‑tuning, and multi‑dimensional evaluation in a single pipeline.

Key technical components

Zero‑code Web UI provides a visual interface for data ingestion, preprocessing, fine‑tuning, and evaluation, eliminating the need for custom scripts.

Automatic data construction uses the KBAlign algorithm to automatically slice domain documents, generate annotation files, and prepare training data for both standard supervised fine‑tuning (SFT) and parameter‑efficient LoRA fine‑tuning.

RAG‑DDR retrieval‑fine‑tuning strategy jointly optimizes dense retrieval and generation models, applying learned retrieval weights during fine‑tuning to improve downstream answer quality.

RAGEval multi‑dimensional evaluation extends ROUGE‑L with metrics for effective information recall and knowledge‑point coverage, enabling robust comparison of different retrieval or generation configurations.

VisRAG multimodal parser extracts tables, images, and formulas from PDF documents, converts them into structured knowledge slices, and stores visual embeddings alongside text embeddings for multimodal retrieval.

Modular architecture separates the system into module, process, and function layers, allowing developers to swap out retrievers, encoders, or LLM back‑ends without modifying the rest of the code base.

Deployment options include a single‑command Docker‑Compose setup and a Conda‑based environment for custom development.

Deployment

Docker (recommended)

# Clone repository (optional)
git clone https://github.com/OpenBMB/UltraRAG.git
cd UltraRAG
# Build and start containers
docker-compose up --build -d
# Access the Web UI at http://localhost:8843

Conda environment

# Create 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
# UI available at http://localhost:8843

Example use case: Legal question‑answering

A corpus of 880 000 regulation excerpts was processed with KBAlign and the RAG‑DDR strategy. Automatic slicing and annotation reduced manual effort by >90 %. After LoRA fine‑tuning, the RAGEval suite reported ROUGE‑L improving from 40.75 % to 53.14 % and a ~30 % increase in key‑clause recall.

Example use case: Multimodal medical knowledge base

VisRAG parsed PDF reports containing textual findings and CT‑image screenshots, extracted visual features, and merged them with text embeddings into unified knowledge slices. Multimodal retrieval returned both relevant narrative excerpts and corresponding image evidence for clinical queries.

Research workflow

Researchers can replace the default retriever component with a custom algorithm, run the built‑in RAGEval evaluation on either the provided benchmark or a user‑supplied dataset, and compare effective information recall and knowledge‑point coverage without re‑implementing data pipelines.

License and source

Apache‑2.0 licensed. Repository: https://github.com/OpenBMB/UltraRAG