Jiuqi Wei

Jiuqi Wei, Quanqing Xu, Chuanhui Yang

Modern AI and vector search are rapidly converging, forming a promising research frontier in intelligent information systems. On one hand, advances in AI have substantially improved the semantic accuracy and efficiency of vector search, including learned indexing structures, adaptive pruning strategies, and automated parameter tuning. On the other hand, powerful vector search techniques have enabled new AI paradigms, notably Retrieval-Augmented Generation (RAG), which effectively mitigates challenges in Large Language Models (LLMs) like knowledge staleness and hallucinations. This mutual reinforcement establishes a virtuous cycle where AI injects intelligence and adaptive optimization into vector search, while vector search, in turn, expands AI's capabilities in knowledge integration and context-aware generation. This tutorial provides a comprehensive overview of recent research and advancements at this intersection. We begin by discussing the foundational background and motivations for integrating vector search and AI. Subsequently, we explore how AI empowers vector search (AI4VS) across each step of the vector search pipeline. We then investigate how vector search empowers AI (VS4AI), with a particular focus on RAG frameworks that integrate dynamic, external knowledge sources into the generative process of LLMs. Furthermore, we analyze end-to-end co-optimization strategies that fully unlock the potential of the ``virtuous cycle" between vector search and AI. Finally, we highlight key challenges and future research opportunities in this emerging area. This paper was published in ICDE 2026.

Chuxu Song, Zhencan Peng, Jiuqi Wei et al.

Long-context LLMs increasingly rely on extended, reusable prefill prompts for agents and domain Q&A, pushing attention and KV-cache to become the dominant decode-time bottlenecks. While sparse attention reduces computation and transfer costs, it often struggles to maintain accuracy at high sparsity levels due to the inherent distribution shift between Queries and Keys. We propose Centroid-Scoring Attention (CSAttention), a training-free sparse attention method optimized for high-throughput serving of reusable contexts. CSAttention adopts a storage-for-computation strategy tailored to the offline-prefill/online-decode setting: it front-loads computation into a one-time offline prefill phase that can be amortized across multiple queries, while aggressively optimizing per-step decoding latency. Specifically, CSAttention constructs query-centric lookup tables during offline prefill, whose size remains fixed during decoding, and enables online decoding to replace full-context scans with efficient table lookups and GPU-friendly score accumulation. Extensive experiments demonstrate that CSAttention achieves near-identical accuracy to full attention. Under high sparsity (95%) and long-context settings (32K-128K), CSAttention consistently outperforms state-of-the-art sparse attention methods in both model accuracy and inference speed, achieving up to 4.6x inference speedup over the most accurate baseline at a context length of 128K.