Qdrant has unveiled version 1.19.0 of its database platform, introducing a TurboQuant storage format that slashes vector storage requirements by up to nine times compared to previous quantization methods. The release, announced by Qdrant, centers on a new Turbo4 datatype that compresses vectors using 4-bit TurboQuant compression as the sole representation, eliminating the need to maintain a full-precision copy. The update delivers both storage savings and throughput gains by reducing the volume of data reads and writes per operation.
The new Turbo4 datatype compresses vectors to four bits without preserving their original full-precision representation, achieving up to nine times less storage compared to TurboQuant quantization alone. This same compression technique applies to multi-vector collections used for ColBERT-style late interaction search, where the benefit scales proportionally larger. Version 1.19 also introduces a unified memory parameter that consolidates per-component memory tier placement across three tiers: pinned, which loads the component entirely into memory where it's never evicted; cached, which keeps data on disk while pre-populating the OS disk cache at startup for fast initial reads that remain evictable under memory pressure; and cold, which loads data lazily from disk on first access. The release adds filtering enhancements including prefix matching on keyword fields and a new slice filter condition for partitioning a collection's points into deterministic, disjoint subsets, plus web UI improvements like live resharding progress and an overhauled Collection Visualizer that scales to tens of thousands of points.
According to the announcement, keyword indexes store values verbatim for exact matching, making them suitable for identifiers like URLs, file paths, and SKUs, but filtering by prefix over these values wasn't possible without either a full payload scan or switching to a text index that tokenizes values and breaks exact matching. The release adds support for prefix matching to keyword indexes by enabling it with "prefix": true in the keyword index configuration, then using the prefix condition in filters. The report states that HNSW graph links can now be pinned in memory, sparse indexes have gained a new cached tier, and quantized vectors can now be pinned, cached, or cold independently of the original vectors' placement.
The unified memory parameter replaces the previous fragmented configuration approach where each component—vectors, the HNSW index, quantized vectors, the sparse index, payloads, and payload indexes—had its own method to configure whether data is loaded into RAM or served from disk using on_disk, always_ram, and on_disk_payload parameters. By consolidating these into a single parameter that works the same way on every component, the release gives users one consistent method to configure the memory tier for any part of a collection, though existing per-component flags remain functional but are deprecated. The new per-tenant IDF statistics feature narrows the IDF corpus to a specific tenant so term rarity reflects that tenant's vocabulary rather than the entire dataset, improving BM25 scoring in multi-tenant deployments. The prefix queries are served from a dedicated index structure, making them as fast as any other indexed filter, which matters when working with large-scale collections where scan performance becomes a bottleneck.
Qdrant positions version 1.19 as a step toward more efficient vector database operations, particularly for deployments handling ColBERT-style searches or multi-tenant environments where storage costs and query performance both matter. The compression gains free up infrastructure budget while the memory tier consolidation simplifies operations, two factors that matter most when vector workloads scale beyond proof-of-concept size. For organizations running large-scale vector search, this release's ability to compress without sacrificing throughput could reshape deployment economics. The memory tier model also signals a broader shift toward giving operators finer control over the performance-cost trade-off without forcing them into all-or-nothing configuration choices.