More about HKUST
From Single-Vector to Multi-Vector Retrieval: Toward Efficient and Scalable Indexing Methods
PhD Qualifying Examination
Title: "From Single-Vector to Multi-Vector Retrieval: Toward Efficient and
Scalable Indexing Methods"
by
Mr. Zhoujin TIAN
Abstract:
Multi-vector retrieval has emerged as a powerful paradigm for semantic search,
where queries and data objects are represented by sets of embeddings rather
than single vectors. By preserving fine-grained query--document interactions,
multi-vector models significantly improve retrieval quality across many
applications. However, this increased expressiveness comes with substantially
higher storage overhead, computational cost, and system design complexity. In
particular, set-level similarity functions are often non-metric, require
expensive pairwise interactions, and complicate efficient indexing and
pruning. To address these challenges, recent work has explored diverse
approaches to scalable multi-vector retrieval, including keyword-style
indexing based on token-level pruning, similarity approximation using
efficient surrogates, and native graph-based index structures that directly
operate on vector sets. Each line of work involves distinct trade-offs
between retrieval accuracy, efficiency, and scalability. This survey provides
a comprehensive overview of recent advances in multi-vector search indexing.
We systematically analyze the core design principles behind existing methods,
discuss their strengths and limitations from both algorithmic and system
perspectives, and outline promising directions for future research in
multi-vector indexing and vector search systems.
Date: Tuesday, 3 February 2026
Time: 9:00am - 11:00am
Venue: Room 2132C
Lift 22
Committee Members: Prof. Xiaofang Zhou (Supervisor)
Prof. Raymond Wong (Chairperson)
Prof. Ke Yi