Cosimo Rulli

Token-aware clustering and hierarchical indexing can slash retrieval latency by an order of magnitude without sacrificing accuracy, making multivector retrieval practical at scale.