README.md

metadata

title: NetraEmbed
emoji: 👁️
colorFrom: yellow
colorTo: purple
sdk: gradio
sdk_version: 6.0.2
app_file: app.py
pinned: false
license: mit
short_description: Universal Multilingual Multimodal Document Retrieval

NetraEmbed - Universal Multilingual Multimodal Document Retrieval

This Space demonstrates NetraEmbed and ColNetraEmbed, state-of-the-art multilingual multimodal document retrieval models based on the BiGemma3 and ColGemma3 architectures.

Features

• NetraEmbed (BiGemma3): Single-vector embedding with Matryoshka representation for fast retrieval
• ColNetraEmbed (ColGemma3): Multi-vector embedding with late interaction for high-quality retrieval with attention heatmaps
• ZeroGPU Integration: Efficient dynamic GPU allocation for on-demand model loading
• PDF Document Support: Upload PDFs and perform semantic search across pages
• Side-by-side Comparison: Compare both models simultaneously

Citation

If you use NetraEmbed or ColNetraEmbed in your research, please cite:

@misc{kolavi2025m3druniversalmultilingualmultimodal,
  title={M3DR: Towards Universal Multilingual Multimodal Document Retrieval},
  author={Adithya S Kolavi and Vyoman Jain},
  year={2025},
  eprint={2512.03514},
  archivePrefix={arXiv},
  primaryClass={cs.IR},
  url={https://arxiv.org/abs/2512.03514}
}

Links

• 📄 Paper
• 💻 GitHub
• 🤗 Models on Hugging Face
• 🌐 CognitiveLab Website

Usage

1. Load Model: Select your preferred model (NetraEmbed, ColNetraEmbed, or Both) and click "Load Model"
2. Upload PDF: Upload a PDF document to index
3. Index Document: Click "Index Document" to process and embed the pages
4. Query: Enter your search query and click "Search" to retrieve relevant pages

This Space uses ZeroGPU for dynamic GPU allocation. Models are loaded on-demand when functions are called.

Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference