Generative Late-Interaction Embeddings for Visual Document Retrieval

Late-interaction search provides state-of-the-art accuracy for visual document retrieval at the cost of storing around 1,000 embeddings per page. Examining this representation, the authors show that page vectors lie on the unit sphere within a low-dimensional manifold of just five to six intrinsic dimensions. Leveraging this geometry, Generative Late-Interaction Embeddings (GLIE) indexes pages using only a few basis vectors and reconstructs the full embedding set on demand via a compact 415K-parameter decoder. On ViDoRe v1, indexing just four vectors per page preserves nearly 80% of uncompressed retrieval performance.