supervc-stack/VectorChord

What it solves

VectorChord addresses the high cost and resource intensity of hosting billion-scale vector databases. It enables users to store and search hundreds of millions to billions of vectors on relatively modest hardware, significantly reducing infrastructure expenses compared to other vector search solutions.

How it works

It is implemented as a PostgreSQL extension that optimizes vector storage and retrieval through several key techniques:

• RaBitQ Compression: Uses native 4-bit (RaBitQ4) and 8-bit (RaBitQ8) vector types to drastically reduce memory and storage footprints while maintaining high search precision.
• Autonomous Reranking: Combines compression with reranking to preserve search quality.
• Hierarchical K-means: Accelerates index building, allowing 100 million vectors to be indexed in approximately 20 minutes.
• Dimensionality Reduction and Sampling: Controls memory growth during the indexing process, enabling the creation of billion-vector indexes on machines with 128GB of RAM.

Who it’s for

Developers and organizations needing to scale vector search to hundreds of millions or billions of embeddings without incurring massive cloud infrastructure costs, particularly those already using or wanting to integrate with PostgreSQL.

Highlights

• Cost Efficiency: Capable of hosting 100M vectors on a single AWS i4i.xlarge instance.
• PostgreSQL Integration: Fully compatible with pgvector data types and syntax for seamless migration.
• High-Speed Indexing: Rapidly builds indexes for massive datasets using optimized disk operations and hierarchical K-means.
• Low-Bit Quantization: Offers native 4-bit and 8-bit types to minimize space without significant recall loss.