Tencent Releases EVIE Series for Visual Document Retrieval

Tencent has released EVIE-8B and its lightweight model EVIE-4.5B, designed to maximize Visual Document Retrieval performance. Both models demonstrated industry-leading accuracy on the ViDoRe V3 benchmark, recording nDCG@10 scores of 66.75 and 66.02, respectively.

Key Technologies and Performance

• High-Capacity Representation: EVIE-8B uses 4096D multi-vector embeddings to preserve fine-grained visual information such as layout, typography, charts, and table structures.
• Prefix-MRL Flexibility: EVIE-4.5B utilizes a single 2048D linear projection to freely truncate dimensions between 64~2048D at runtime, allowing it to handle diverse retrieval requirements without separate models.
• Ultra-Compact Index: By applying Hierarchical Agglomerative Clustering, the number of vectors per page was compressed from approximately 750 to 32. This significantly reduced index storage space to 3.81 GiB per million pages.

Validation and Distillation Techniques

Both models were validated across 138 multilingual tasks, including ViDoRe V1, V2, V3, and JinaVDR, using four metrics: nDCG, Recall, MAP, and MRR. EVIE-4.5B applied the EVIE-ARD distillation recipe with EVIE-8B as the teacher model, reproducing performance similar to full training despite being lightweight.