Tencent Releases EVIE-8B and EVIE-4.5B Models for Visual Document Retrieval
Tencent has open-sourced EVIE-8B and EVIE-4.5B, state-of-the-art vision-language models for visual document retrieval that set a new benchmark high score of 66.75 nDCG@10 on ViDoRe V3. The models utilize full per-token multi-vector representations alongside Prefix-Matryoshka Representation Learning to provide high retrieval accuracy with dynamic vector sizes. These models advance multimodal Retrieval-Augmented Generation (RAG) by allowing systems to accurately search visually complex documents containing tables, charts, and custom layouts. By leveraging Matryoshka embeddings and token clustering, EVIE drastically cuts index storage to 3.81 GiB per million pages while maintaining flexible compute requirements. EVIE-8B employs 4096-dimensional multi-vector embeddings per token and serves as a teacher model to train EVIE-4.5B via anchor-preserving relation distillation. EVIE-4.5B uses Hierarchical Agglomerative Clustering (HAC) to compress roughly 750 token vectors per page down to 32 vectors, allowing runtime dimension truncation between 64 and 2048 dimensions.
## BACKGROUND
Visual Document Retrieval (VDR) evaluates how effectively AI models search and retrieve relevant context from visually complex PDFs, charts, and scanned pages, as measured by benchmarks like ViDoRe. Matryoshka Representation Learning (MRL) is a technique that encodes information at nested granularities within a single embedding vector, enabling runtime truncation to reduce storage and memory usage without re-training.