Qdrant 1.19's Turbo4 Datatype Cuts Vector Storage 9x by Deleting the Full-Precision Copy — and Openly Admits You Lose Rescoring

Qdrant Blog·low signalXBlueskyLinkedInCopy link

Released August 5, Qdrant 1.19 promotes last version's TurboQuant from a secondary quantization layer to a primary storage datatype: Turbo4 keeps only the 4-bit representation, dropping from 36 bits per coordinate (float32 original plus 4-bit copy) to 4 bits, a ninefold reduction that also cuts per-operation disk reads and writes. The post is unusually candid about the tradeoff — without a full-precision copy Qdrant cannot rescore top candidates, so Turbo4 is for disk-bound deployments while TurboQuant-over-full-precision remains correct when recall is the priority; the gain is proportionally larger for ColBERT-style multi-vector collections. The same release adds per-component memory tiers (`cold`/`cached`/`pinned`), prefix matching in keyword filters, per-query IDF for sparse search, deterministic sliced scroll/sampling, and a global quota API.