Optimizing Visual Document Retrieval with Late-Interaction Embeddings
September 9, 2026
Visual document retrieval accuracy improves by normalizing k-means centroids to the unit sphere, correcting for systematic MaxSim underestimation. This geometric correction provides up to +0.093 nDCG@5 over raw centroids for compressed late-interaction embeddings.
HOW THIS AFFECTS YOU
●
builderYou can improve retrieval accuracy for document search engines with minimal storage overhead.
●
researcherThis identifies specific geometric properties in vector embeddings that affect retrieval performance.