# VideoRAG
**Repository Path**: turbig/VideoRAG
## Basic Information
- **Project Name**: VideoRAG
- **Description**: [KDD'2026] "VideoRAG: Chat with Your Videos"
arxiv.org/abs/2502.01549
- **Primary Language**: Unknown
- **License**: Not specified
- **Default Branch**: main
- **Homepage**: None
- **GVP Project**: No
## Statistics
- **Stars**: 0
- **Forks**: 0
- **Created**: 2026-08-05
- **Last Updated**: 2026-08-05
## Categories & Tags
**Categories**: Uncategorized
**Tags**: None
## README
Vimo is a revolutionary desktop application that lets you **chat with your videos** using cutting-edge AI technology. Built on the powerful [VideoRAG framework](https://arxiv.org/abs/2502.01549), Vimo can understand and analyze videos of any length - from short clips to hundreds of hours of content - and answer your questions with remarkable accuracy.
### ๐ฅ Watch Vimo in Action
See how Vimo transforms video interaction with intelligent conversations and deep understanding capabilities.
๐ Click to watch the Vimo demo video
## โจ Key Features
### For Everyone
- **Drag & Drop Upload**: Simply drag video files into Vimo
- **Smart Conversations**: Ask questions in natural language
- **Multi-Format Support**: Works with MP4, MKV, AVI, and more
- **Cross-Platform**: Available on macOS, Windows, and Linux
### For Power Users
- **Extreme Long Videos**: Process videos up to hundreds of hours
- **Multi-Video Analysis**: Compare and analyze multiple videos simultaneously
- **Advanced Retrieval**: Find specific moments and scenes with precision
- **Export Capabilities**: Save insights and references for later use
### For Researchers
- **VideoRAG Framework**: Access to cutting-edge retrieval-augmented generation
- **Benchmark Dataset**: LongerVideos benchmark with 134+ hours of content
- **Performance Metrics**: Detailed evaluation against existing methods
- **Extensible Architecture**: Build upon our open-source foundation
## ๐ Why Vimo?
**For Video Enthusiasts & Professionals:**
- **Effortless Video Analysis**: Upload any video and start asking questions immediately
- **Natural Conversations**: Chat with your videos as if talking to a human expert
- **No Length Limits**: Process everything from 30-second clips to 100+ hour documentaries
- **Deep Understanding**: Combines visual content, audio, and context for comprehensive answers
**For Researchers & Developers:**
- **State-of-the-Art Algorithm**: Built on VideoRAG, featuring graph-driven knowledge indexing
- **Benchmark Performance**: Evaluated on 134+ hours across lectures, documentaries, and entertainment
- **Open Source**: Full access to VideoRAG implementation and research findings
- **Scalable Architecture**: Efficient processing with single GPU (RTX 3090) capability
## ๐ Table of Contents
- [๐ Quick Start](#-quick-start)
- [โจ Key Features](#-key-features)
- [๐ฌ VideoRAG Algorithm](#-videorag-algorithm)
- [๐ ๏ธ Development Setup](#๏ธ-development-setup)
- [๐งช Benchmarks & Evaluation](#-benchmarks--evaluation)
- [๐ Citation](#-citation)
- [๐ค Contributing](#-contributing)
- [๐ Acknowledgement](#-acknowledgement)
## ๐ Quick Start of Vimo
### Option 1: Download Vimo App (Coming Soon)
> [!NOTE]
> We are preparing the **Beta release** for macOS Apple Silicon first, with Windows and Linux versions coming soon!
### Option 2: Run from Source Code
For detailed setup instructions:
- **Vimo Desktop App**: See [Vimo-desktop](Vimo-desktop) for complete installation and configuration steps
**Quick Overview:**
1. Set up the Python backend environment and start the VideoRAG server
2. Launch the Electron frontend application
3. Start chatting with your videos!
## ๐ฌ VideoRAG Algorithm
VideoRAG introduces a novel dual-channel architecture that combines:
- **Graph-Driven Knowledge Indexing**: Multi-modal knowledge graphs for structured video understanding
- **Hierarchical Context Encoding**: Preserves spatiotemporal visual patterns across long sequences
- **Adaptive Retrieval**: Dynamic retrieval mechanisms optimized for video content
- **Cross-Video Understanding**: Semantic relationship modeling across multiple videos
### Technical Highlights
- **Efficient Processing**: Handle hundreds of hours on a single RTX 3090 (24GB)
- **Structured Indexing**: Distill long videos into concise knowledge representations
- **Multi-Modal Retrieval**: Align textual queries with visual and audio content
- **LongerVideos Benchmark**: 160+ videos, 134+ hours across diverse domains
### Performance Comparison
Our VideoRAG algorithm significantly outperforms existing methods in long-context video understanding:
We also evaluate VideoRAG's QA performance on the Video-MME long video track to better understand the gains over the backbone models (included here because of the paper's page limit):
| Video-MME Long Video | MiniCPM-o w/o subs | MiniCPM-o w/ subs | MiniCPM-V w/o subs | MiniCPM-V w/ subs | VideoRAG |
| --- | ---: | ---: | ---: | ---: | ---: |
| Accuracy | 52.2% | 56.3% | 51.8% | 56.3% | **60.2%** |
> Note: The score may show slight fluctuations across runs due to the instability of LLM generation.
### Experiments and Evaluation
See [VideoRAG-algorithm](VideoRAG-algorithm) for detailed development setup including:
- Conda environment creation
- Model checkpoints download
- Dependencies installation
- Evaluation scripts
## ๐งช LongerVideos Benchmark
We created the LongerVideos benchmark to evaluate long-context video understanding:
| Video Type | #Collections | #Videos | #Queries | Avg. Duration |
|------------------|-------------|---------|----------|---------------|
| **Lectures** | 12 | 135 | 376 | ~64.3 hours |
| **Documentaries**| 5 | 12 | 114 | ~28.5 hours |
| **Entertainment**| 5 | 17 | 112 | ~41.9 hours |
| **Total** | 22 | 164 | 602 | ~134.6 hours |
For detailed evaluation instructions and reproduction scripts, see [VideoRAG-algorithm/reproduce](VideoRAG-algorithm/reproduce).
## ๐ Citation
If you find Vimo or VideoRAG helpful in your research, please cite our paper:
```bibtex
@article{VideoRAG,
title={VideoRAG: Retrieval-Augmented Generation with Extreme Long-Context Videos},
author={Ren, Xubin and Xu, Lingrui and Xia, Long and Wang, Shuaiqiang and Yin, Dawei and Huang, Chao},
journal={arXiv preprint arXiv:2502.01549},
year={2025}
}
```
## ๐ค Contributing
We welcome contributions from the community! Whether you're:
- **Reporting bugs** or suggesting features for Vimo
- **Improving VideoRAG algorithms** or adding new capabilities
- **Enhancing documentation** or creating tutorials
- **Designing UI/UX improvements** for better user experience
Feel free to submit issues and pull requests. Together, we're building the future of intelligent video interaction!
## ๐ Acknowledgement
Vimo builds upon the incredible work of the open-source community:
- **[VideoRAG](https://arxiv.org/abs/2502.01549)**: The core algorithm powering Vimo's intelligence
- **[nano-graphrag](https://github.com/gusye1234/nano-graphrag)** & **[LightRAG](https://github.com/HKUDS/LightRAG)**: Graph-based retrieval foundations
- **[ImageBind](https://github.com/facebookresearch/ImageBind)**: Multi-modal representation learning
- **[uitars-desktop](https://github.com/bytedance/UI-TARS-desktop)**: Desktop application architecture inspiration
**๐ Transform how you interact with videos. Start your journey with Vimo today!**
---
Built with โค๏ธ by the VideoRAG@HKUDS team.
[](https://learnopencv.com/videorag-long-context-video-comprehension/)
[]()
**๐ฌ Intelligent Video Conversations | Powered by Advanced AI | Extreme Long-Context Processing**
