MarginMerge Compression: Preserves 97% Quality for Visual Document Retrievers with 90% Fewer Vectors

_reachsumit · x · 2026-08-05

Multi-vector visual document retrievers like ColPali produce large indexes and costly scoring due to storing fine-grained patch embeddings.

This paper introduces MarginMerge, a compression method that selects coverage-aware anchors, clusters document patches, and synthesizes representative vectors. Experiments show that while reducing stored document vectors by 90% to 95%, this method preserves 97% to 99% of the average nDCG@5 retrieval quality compared to the uncompressed index.

Original post →

More from Research

Research channel →