일시: 2026. 7. 3. (금요일), 오후 2시

장소: 제4공학관 D601호

Speaker: Yinjun Wu(Ph.D.), Assistant Professor of Computer Science at Peking University

Title: Multi-Vector Retrieval for LLM Systems: A Coherent System Stack and Future Extensions

Abstract:

Vector search has become a cornerstone of modern LLM systems, enabling retrieval-augmented generation (RAG) and semantic memory. Multi-vector retrieval (MVR) is an emerging paradigm within vector search that achieves superior accuracy by representing each query and document as a set of fine-grained embeddings. This talk presents a coherent system stack built on MVR as a first-class primitive for LLM-powered applications, highlighting the fundamental differences from systems built on single-vector retrieval. The stack integrates three layers: POQD, a learned query parser that optimizes natural language decomposition into fine-grained segments for MVR; MVR-cache, a semantic cache layer that uses multi-vector representations to accurately identify semantically equivalent prompts, thereby eliminating unnecessary LLM invocations; and Multi-HNSW, a lightweight graph-based index that constructs an HNSW-style graph-based index over all individual document vectors, enabling efficient document-aware retrieval.

Beyond these core components, I will discuss ongoing efforts that leverage the same MVR infrastructure for LLM agent memory management, particularly in the context of long-context memory, and using vector search as a surrogate to accelerate semantic query processing for cost reduction over large-scale databases. Together, this stack and its extensions illustrate how MVR can serve as a unifying primitive for retrieval, caching, memory, and query processing in LLM systems.

Bio:

Yinjun Wu is an assistant professor of computer science at Peking University starting from 2024. He received his PhD from the University of Pennsylvania in 2021. His current research interests lie at the intersection of data science, data management, and machine learning, with a focus on making data systems more efficient, robust, and queryable. He has published in top-tier venues including VLDB, SIGMOD, ICDE, ICML, and AAAI, covering topics such as data cleaning, query optimizations, causality for data systems, and vector search.