GLIE aims to cut vector storage for visual document search, a preprint announcement says
The announcement says late-interaction retrievers store roughly 1,000 vectors per page, but GLIE could use a handful plus a small shared decoder to reconstruct the full set.
TLDR
A post introducing GLIE—Generative Late-Interaction Embeddings for Visual Document Retrieval—describes a two-stage search process. A handful of stored vectors, numerical representations of each page, serve as the search index. According to the announcement, a small shared decoder then regenerates the full vector set only for the top candidates so they can be reranked.
Combined views
—
1 Source, first seen ago
