Binary Quantization – Now Available for the Neo4j Vector Index

Session track: Modern Applications

Session time: 08:45 am - 09:00 am UTC November 12

Session description:

Searching a Vector Index can be a costly operations, and with an increasing number of vector properties in the index, the memory required to perform a search also increases and quickly becomes a bottleneck. To mitigate this, quantization can be used to reduce the fidelity of vectors for a search, reducing the memory footprint at the cost of a loss in accuracy. Neo4j had offered Scalar Quantization for some time, and starting with 2026.06 as a preview feature, users can now use Binary Quantization, which only uses a single bit for each vector, as well as automatic rescoring via over-fetching. In this session, Paul will give you an overview of quantization, why and when you would want to use it. You'll learn how Scalar and, in particular, Binary Quantization works, how it can be used in Neo4j, and how that impacts resource footprint and result accuracy. You'll also learn about rescoring, and how it can be used to improve the result accuracy. This session will go beyond the usage and Paul will also explain how Binary Quantization can achieve usable accuracy while reducing vector fidelity down to just a single bit.