EVIE:证据向量驱动的视觉文档检索嵌入
原标题:EVIE: Evidence-Vector-Informed Embeddings for Visual Document Retrieval
AI 摘要
该论文提出 EVIE(Evidence-Vector-Informed Embeddings),一种原生视觉文档检索器系列,通过证据判定的数据治理、双向师生蒸馏与 Prefix-MRL、以及层次凝聚索引压缩(HAC)三项创新,解决视觉文档检索中细粒度匹配与索引存储的矛盾。EVIE-8B 在 ViDoRe V3 上取得 66.75 nDCG@10,超过最佳外部基线 1.43 分,四套件平均 79.51;EVIE-4.5B 配合 HAC 在每百万页仅 3.81 GiB 存储下保留 59.58 nDCG@10。实验覆盖 ViDoRe V1、V2、V3 和 JinaVDR 共 138 项任务,改善了视觉文档检索的精度-存储权衡。
正文节选
EVIE: Evidence-Vector-Informed Embeddings for Visual Document Retrieval Abstract Accurate and scalable visual document retrieval (VDR) requires both fine-grained page understanding and efficient indexing, yet existing approaches struggle to achieve both. OCR-based text retrieval adds preprocessing latency and can lose visual and structural cues needed to understand complex pages. Single-vector vision-language models bypass OCR, but compressing an entire page into one vector limits the granularit