Qdrant has released version 1.19 of its vector database, combining a new four-bit vector format with broader changes to memory management, filtering, administration and distributed deployments. The update is substantial for teams trying to fit larger retrieval collections onto fixed infrastructure.
The headline feature is Turbo4, a datatype that stores vectors using four-bit TurboQuant encoding without retaining a full-precision copy. Qdrant says this can cut vector storage by as much as nine times compared with conventional full-precision storage.
Smaller vectors come with an explicit trade-off
Turbo4 is intended for collections where memory or disk cost is a stronger constraint than maximum recall. Because the full-precision vector is not available for exact rescoring, the compressed representation can lose some retrieval quality. Qdrant recommends rescoring strategies where an alternative representation is available, but teams should benchmark against their own queries before converting a production collection.
The practical gain could be meaningful for large recommendation or retrieval-augmented generation systems. More vectors may fit in memory, or the same collection may run on less expensive hardware. The right metric is not compression alone: teams need to compare recall, answer quality, latency and infrastructure cost as a set.
One memory policy replaces scattered switches
Version 1.19 adds a unified memory parameter with pinned, cached and cold modes across components. Older component-specific flags are deprecated. The consolidated control should make resource behaviour easier to reason about, particularly in deployments where vector data, payloads and indexes previously had separate configuration paths.
Pinned data is kept ready in memory, cached data can use the operating system’s page cache, and cold data favours storage over memory residency. Operators can choose policies according to access patterns rather than applying the same treatment to an entire collection.
As with any resource-management change, migration deserves observation under normal and peak traffic. Cache warm-up, storage speed and working-set size can produce results that do not appear in a small pre-production test.
Filtering gets closer to application language
The release adds per-tenant inverse document frequency, helping text relevance account for the vocabulary of an individual tenant rather than the full shared collection. It also introduces prefix matching for keyword fields and slice filters, providing more precise ways to narrow retrieval before vector scoring.
Qdrant has made BM25 language-neutral, which simplifies text-search configuration across multilingual content. Faster faceting should improve workloads that need counts or category summaries alongside results.
Operational changes reach the dashboard and cluster
The web interface now shows resharding progress and uses a server-side collection visualiser designed to handle tens of thousands of points with WebGL2 and WebAssembly. Payload indexes can also be managed from the interface. These changes make it easier to inspect a collection without pulling excessive data into the browser.
For larger installations, Qdrant adds resource quotas and read affinity. Quotas can limit the resources consumed by a tenant or workload, while read affinity lets operators favour replicas that better match a routing policy.
The release removes the legacy search, recommend and discover endpoints in favour of the unified /query endpoint. That is the compatibility change to identify before upgrading. Qdrant Cloud handles upgrades automatically; self-hosted deployments should move one version at a time, including intermediate patch versions where required, and verify that client libraries no longer depend on removed routes.
Qdrant 1.19 offers real efficiency gains, but its most valuable settings are workload-dependent. A controlled benchmark and staged rollout will reveal whether Turbo4’s smaller footprint and the new memory policy improve the application without weakening the retrieval quality users actually see.