ColNanoVDR enables document-free distillation of multi-vector retrievers by aligning student query tokens with teacher query embeddings using Optimal Transport. This avoids the need to encode and cache terabytes of training page tokens.
HOW THIS AFFECTS YOU
●
builderYou can train efficient visual document retrievers without the massive storage costs of page-level tokens.
●
researcherThis introduces a new objective, OTW, for aligning multi-vector student models to teachers.