UltraRAG: A One-Stop RAG System Solution to Simplify Data Construction and Model Fine-Tuning

UltraRAG is a RAG (Retrieval Augmented Generation) system solution jointly proposed by the THUNLP group of Tsinghua University, NEUIR group of Northeastern University, Modelbest.Inc and 9#AISoft team. Based on agile deployment and modular construction, the framework provides an automated system of data construction, model fine-tuning, and inference evaluation techniques.UltraRAG significantly simplifies the whole process from data construction to model fine-tuning, and helps researchers and developers to deal with complex tasks efficiently. Its code-free programming WebUI supports users to easily manipulate the full chain of setup and optimization processes, including the multimodal RAG solution VisRAG.

Function List

• Code-free programming WebUI support: Users can operate the full link setup and optimization process without programming experience.
• One-click synthesis and fine-tuning solutions: Based on proprietary methods such as KBAlign and RAG-DDR, the system supports one-click systematic data construction and retrieval, and performance optimization through diverse model fine-tuning strategies.
• Multi-dimensional, multi-stage robust assessmentThe core RAGEval methodology, combined with a multi-stage evaluation approach, significantly enhances the robustness of the "model evaluation".
• Research Friendly Explore Work Integration: Includes the THUNLP-RAG group's proprietary methodology and other cutting-edge RAG methods to support ongoing module-level exploration and development.
• Rapid deployment: Supports rapid deployment via Docker and Conda, making it easy for users to get started quickly.

Using Help

environmental dependency

• CUDA version 12.2 or above is required.
• Python version needs to be 3.10 or above.

Rapid deployment

Deployment via Docker

1. Run the following command:

   docker-compose up --build -d

2. Access in browserhttp://localhost:8843The

Deployment via Conda

1. Create the Conda environment:

   conda create -n ultrarag python=3.10

2. Activate the Conda environment:

   conda activate ultrarag

3. Install relevant dependencies:

   pip install -r requirements.txt

4. Run the following script to download the model (by default it downloads to theresources/models(Catalog):

   python scripts/download_models.py

5. Run the demo page:

   streamlit run ultrarag/webui/webui.py --server.fileWatcherType none

Main function operation flow

Programming WebUI without code

1. Visit the WebUI page and select the desired RAG solution (e.g. VisRAG).
2. Setup for data construction, model fine-tuning, and inference evaluation based on prompts.
3. By clicking the "One-click Synthesis and Fine-tuning" button, the system will automatically complete the data construction and model fine-tuning.

Multi-dimensional, multi-stage robust assessment

1. Select the RAGEval assessment method in the WebUI.
2. Set the evaluation parameters and click the "Start Evaluation" button.
3. The system will automatically perform a multi-stage assessment and generate an assessment report.

Research Friendly Explore Work Integration

1. Select the desired RAG method in the WebUI (e.g. THUNLP-RAG).
2. Follow prompts for module-level exploration and development.
3. Click on the "Start Exploring" button and the system will automatically explore and develop.