Cheng LONG
I work on vector databases, making similarity search over high-dimensional embeddings fast, memory-efficient, and accurate. I lead VectorDB @ NTU.
My group developed RaBitQ, a vector quantization method with theoretical error guarantees, now used in Faiss, Milvus, NVIDIA cuVS, and other production systems. My broader interests include machine learning and data management for spatial data, time series, and graphs.
RaBitQ: vector quantization with guarantees
RaBitQ combines compact vector codes with an unbiased distance estimator and a provable error bound. The original method uses one bit per dimension; its multi-bit extension reaches an asymptotically optimal space–error trade-off. Split-RaBitQ supports compression ratios beyond 32×. The 32× reference compares one-bit vector codes with float32 vectors, excluding auxiliary factors and index overhead.
- RemoteANNarXiv 2026 IVF-RaBitQ for vectors on cloud object storage.
- RaBitQ-SSDarXiv 2026 Split codes and pipelined I/O for SSD-resident vectors.
- IVF-RaBitQ on GPUVLDB 2026 GPU-native index build and search, integrated into NVIDIA cuVS.
Foundations and earlier developments
- RaBitQ LibraryVecDB @ ICML 2025 Open-source C++ library with IVF and graph indexes.
- SymphonyQGSIGMOD 2025 Combines RaBitQ with graph indexes.
- Extended RaBitQSIGMOD 2025 Multiple bits per dimension; practical and asymptotically optimal.
- RaBitQSIGMOD 2024 1 bit per dimension, with a theoretical error bound for ANN search.
Integrating the RaBitQ Library: txtai · Zvec
RaBitQ across the vector-search ecosystem: Milvus · Faiss · NVIDIA cuVS · Microsoft DiskANN · VSAG · VectorChord · Volcengine OpenSearch · CockroachDB · turbopuffer · Zvec · LanceDB · Databricks · ClickHouse · Qdrant · Weaviate · pg_pinecone · Elasticsearch and Apache Lucene (RaBitQ-inspired binary quantization, BBQ)
Other research
More on vector search
Filtered and hybrid search, out-of-distribution queries, and fast distance comparisons.
Spatial data, time series & graphs
LLMs for time series and spatio-temporal prediction, and learned indexes built with reinforcement learning.
News
- Oct 2026RemoteANN and RaBitQ-SSD posted on arXiv.
- Jun 2026GPU-native IVF-RaBitQ accepted to VLDB 2026.
- Jan 2026Received the CCDS Research Award (Young Faculty).
- Oct 2025Jianyang GAO receives the CCDS Outstanding PhD Thesis Award.
Selected talks
- From Billions of Embeddings to Trillions of Weights: A Quantization Story Alibaba; Global Software Technology Summit, Asia-Pacific (2026)
- High-Dimensional Vector Quantization: Framework, Advances, and Future Directions CCDS Research Fest (2026); Alibaba Cloud (2025)
- Nearest Neighbor Search on High-Dimensional Vector Data Keynote, TheWebConf'24 Workshop
- RL4SpatialDB: Reinforcement Learning for Spatial Data Management Keynote, MUST 2023; Microsoft Research Redmond (2023)
Selected awards
- 2026CCDS Research Award (Young Faculty), NTU
- 2024Best Demo Runner-Up, ICDE
- 2014Fulbright-RGC Hong Kong Research Scholar Award
- 2013Prof. Francis Chin Research Award, ACM Hong Kong
Service
- Senior PC, ICDE 2027
- Workshop Co-Chair, SIGMOD 2026
- Associate Editor, Data & Knowledge Engineering