Clustering Optimization for Vector Embeddings: 1-Bit Codes Achieve Near-Optimal Quality
A recent study published on arXiv reevaluates three popular methods in vector search: dimensionality reduction, quantization, and dimension pruning, aiming to enhance vector embedding indexing through clustering. The researchers introduce an indexing framework that incorporates these methods prior to clustering, examining their effects on storage space, clustering duration, and the quality of centroids in vector search applications. Results indicate that utilizing full-precision vectors for clustering is unnecessary; even 1-bit codes can yield clustering quality close to optimal (within 1% of the ideal), while cutting storage needs by 60 times and improving performance. The implementations are available as open-source. This research is particularly significant for the digital art and AI technology industries, where efficient vector search is vital for image retrieval and recommendation systems.
Key facts
- The study revisits dimensionality reduction, quantization, and dimension pruning for vector search.
- The proposed pipeline applies these techniques before clustering.
- Even 1-bit codes achieve near-optimal clustering quality (within 1% of ideal).
- Storage requirements are reduced by 60x with 1-bit codes.
- The implementations are open-sourced.
- The study is from arXiv (Computer Science > Databases).
- The paper is titled 'Stop Indexing at Full Precision: Revisiting Clustering for Vector Embeddings'.
- The paper was submitted to arXiv with ID 2608.14648.
Entities
Institutions
- arXiv