Multilingual Visual Document Retrieval Model Released

vdr-2b-multi-v1 is an embedding model designed to encode document page screenshots into dense vectors, enabling retrieval of visually rich multilingual documents without OCR or data extraction processes. It was developed in collaboration with LlamaIndex and delivers stronger performance than existing models.

Key features are as follows:

• Multilingual support: Supports Italian, Spanish, English, French, and German, and to this end, a 500,000-item high-quality multilingual synthetic dataset (vdr-multilingual-train) was released alongside the model.
• Efficient inference: Provides 3x faster inference speed and lower VRAM usage compared to existing models.
• Cross-lingual retrieval: Enables cross-lingual retrieval, such as finding German documents using an Italian query.
• Matryoshka Representation Learning: Allows vector size to be reduced by 3x while retaining 98% of quality, optimizing storage costs and retrieval speed.

This model integrates directly with LlamaIndex and SentenceTransformers, making it easy to use.