VectorHit
Ongoing research on GPU-accelerated multi-vector retrieval.
GPU-Accelerated, Multi-Vector Retrieval
Ongoing research
Multi-vector retrieval represents queries and documents with sets of vectors, enabling fine-grained matching for text and multimodal search. MaxSim/Chamfer-style scores aggregate the best document-vector match for each query vector. This improves expressiveness but makes candidate discovery and scoring computationally expensive.
VectorHit explores GPU-efficient candidate generation and scoring for these workloads, with the goal of supporting high-quality retrieval at scale.
Topics: vector search · late interaction · MaxSim · multimodal retrieval · GPU systems