Sai Huang

Sai Huang

Auto-generated profile, based on publicly available information

Sai Huang

Alibaba Cloud, Hangzhou, China

Research Focus & Metrics

Academic Career Length

1Year (2026-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
30 May 2026
Verified
Conference Paper
Open access
YWYan Wang
Yan Wang
JZJian Zhou
Jian Zhou
SHSai Huang
Sai Huang
CDChao Dou
Chao Dou
HTHanwen Tian
Hanwen Tian
DDDong Deng
Dong Deng
Vector databases have become a cornerstone of modern AI infrastructure, enabling semantic retrieval and retrieval-augmented generation (RAG) over massive unstructured datasets. However, existing systems face an inherent trade-off: specialized vector databases deliver low-latency in-memory search but struggle with scalability and integration with structured data, whereas general-purpose databases provide consistency and fault tolerance at the cost of query performance. In this paper, we present LindormVector, a distributed, scalable, and cost-efficient vector engine built natively on Lindorm, Alibaba Cloud's multi-model, cloud-native NoSQL database. LindormVector adopts a shared-storage architecture where all data is persisted in a distributed file system for elasticity, durability, and high availability. To minimize I/O operations, network communication, and memory footprint, LindormVector adopts an IVFPQ-style index tailored to Lindorm's storage model. A key distinction in our index design lies in the number of clusters nlist. We observe that the optimal number of clusters is often substantially larger than the conventional sqrt(n) rule-of-thumb for a dataset of n vectors, especially when n is large. To guide the choice of nlist, we develop a practical cost model capturing the trade-off between centroid search and code scan costs. To efficiently support large nlist values, LindormVector accelerates K-Means clustering via approximate nearest neighbor search, organizes centroids with an HNSW graph, and introduces an optimized greedy graph traversal for efficient centroid search. LindormVector further supports real-time index updates and efficient filtered and hybrid vector search. Extensive experiments on billion-scale datasets demonstrate that LindormVector consistently achieves high recall with millisecond-level latency and significantly outperforms both specialized and general-purpose vector databases including Milvus, Elasticsearch, and PGVector. LindormVector is already deployed in production across Alibaba Cloud, where the largest cluster consists of hundreds of ECS instances and manages more than 100 billion vectors.