多模態檢索與 RAG
Tencent’s EVIE Cuts Visual Document Retrieval Index Costs with Truncatable Multi-Vectors and Token Clustering
Tencent has released EVIE-8B and EVIE-4.5B, which create per-token representations directly from page images instead of performing OCR before retrieval. The 4.5B model supports runtime-adjustable vector dimensions and can compress each page to 32 vectors, although its leading results currently come mainly from the team’s own evaluations.
Tencent has released the EVIE-8B and EVIE-4.5B weights on Hugging Face, along with inference and training code and evaluation pipelines covering 138 tasks. Rather than first converting PDFs, tables, or scanned pages into plain text, these visual document retrievers use a Qwen3.5 vision-language backbone to produce per-token multi-vector representations. MaxSim then matches a text query against local visual representations within each page. This allows RAG systems that are sensitive to fonts, layouts, charts, and table positions to avoid compressing all information into a single dense embedding.
The 8.41B-parameter EVIE-8B uses 4,096-dimensional representations and primarily serves as a high-capacity teacher. The more deployment-oriented EVIE-4.5B uses Prefix-MRL to train a single 2,048-dimensional projection that can be truncated at runtime to 64, 128, 256, 512, 1,024, or 2,048 dimensions, eliminating the need to store multiple checkpoints. A separate HAC pipeline, which requires no additional training, clusters visual tokens according to feature and spatial distances. In Tencent’s example, it compresses roughly 750 vectors to 32 vectors per page, reducing the estimated raw BF16 index for one million pages to 3.81 GiB. Engineering teams must still evaluate the vector database structure, MaxSim computation, and retrieval quality together rather than comparing model file sizes alone.
Tencent reports ViDoRe V3 nDCG@10 scores of 66.75 for EVIE-8B and 66.02 for EVIE-4.5B. However, these figures were produced by the publisher using its own evaluation pipeline, and neither a formal paper nor architectural ablation studies have been released. The model card’s own tables also show that EVIE-8B does not rank first on ViDoRe V1. The 8B checkpoint is also the result of merging weights from two training branches. The next questions are whether third parties can reproduce the reported results and whether HAC can preserve recall on Chinese financial reports, low-resolution scans, and large-scale approximate vector indexes.