研究库 论文知识库

内容库 / 论文

arXiv:2609.11808 · RAG 检索增强

Generative Late-Interaction Embeddings For Visual Document Retrieval

面向视觉文档检索的生成式晚交互 embedding

Generative Late-Interaction Embeddings For Visual Document Retrieval

类型：arxiv
标识：2609.11808
链接：https://arxiv.org/abs/2609.11808
主分类：rag
形态：method
TLDR：Late-interaction retrieval is the state-of-the-art for visual document search, but it pays for its accuracy in storage. Existing compression methods retain a subset or local average of the N~1,000 vectors per page. Under aggressive storage budgets, however, these methods degrade sharply, and alternatives require retraining the encoder. Investigating this degradation across three encoders, we found two consistent properties: the vectors lie exactly on the unit sphere and concentrate near a manifold of intrinsic dimension five to six. This geometry yields two insights. First, standard k-means ce
待LLM分类：否
标题中文：面向视觉文档检索的生成式晚交互 embedding
TLDR中文：晚交互检索是当前视觉文档搜索的 SOTA 方法，但其准确性是以存储开销为代价的。现有压缩方法仅保留每页约 N~1,000 个向量中的子集或局部均值。然而在激进的存储预算下，这些方法性能急剧下降，而替代方案又需重新训练编码器。通过在三种编码器上分析这种下降现象，我们发现两个一致特性：向量恰好位于单位球面上，并集中分布在内禀维度为 5 至 6 的流形附近。该几何特性带来两点启示。首先，标准 k-means 中心……
来源文件：
/inbox/tom/_candidates/2026-09-11-agent-rag-longcontext-candidates.json

相关 · 按语义匹配

笔记 2026-07-03 轻量精读：Qwen-Image-Agent × The Verification Horizon

仓库 iron-XiangruChen/reservoir-image-review-reproducibility

笔记精读：Efficient Inference for Large Vision-Language Models: Bottlenecks, Techniques, and Prospects

仓库 griver/Q-RAG

笔记 Cool Papers · cs.IR (papers.cool) · RSS 摘要

笔记 VaLR: Vision-aligned Latent Reasoning for Multi-modal LLM

arXiv

2609.11808

TLDR

晚交互检索是当前视觉文档搜索的 SOTA 方法，但其准确性是以存储开销为代价的。现有压缩方法仅保留每页约 N~1,000 个向量中的子集或局部均值。然而在激进的存储预算下，这些方法性能急剧下降，而替代方案又需重新训练编码器。通过在三种编码器上分析这种下降现象，我们发现两个一致特性：向量恰好位于单位球面上，并集中分布在内禀维度为 5 至 6 的流形附近。该几何特性带来两点启示。首先，标准 k-means 中心……Late-interaction retrieval is the state-of-the-art for visual document search, but it pays for its accuracy in storage. Existing compression methods retain a subset or local average of the N~1,000 vectors per page. Under aggressive storage budgets, however, these methods degrade sharply, and alternatives require retraining the encoder. Investigating this degradation across three encoders, we found two consistent properties: the vectors lie exactly on the unit sphere and concentrate near a manifold of intrinsic dimension five to six. This geometry yields two insights. First, standard k-means ce

相关解读

arXiv ↗ 中文全文