RaBitQ arrives in Vectorium with claims of up to 30x compression
The announcement says the integration needs no training, supports 1, 2, 4 or 8 bits per component, and runs 1.4–2.2x faster than the original implementation.
TLDR
RaBitQ has been added to Vectorium, according to an announcement claiming up to 30x compression without training. The post lists support for 1, 2, 4 or 8 bits per component and claims performance 1.4–2.2x faster than the original implementation. For neural embeddings, it also claims RaBitQ beats PQ (product quantization) on every axis, with the gap growing as dimensionality increases.
Combined views
—
1 Source, first seen ago
