In brief

Qdrant 1.19, released on August 5, 2026, allows the original vectors in a collection to be stored directly in 4-bit TurboQuant format. For RAG systems, that can materially reduce the disk footprint: for the raw vector array, the theoretical difference between `float32` and four bits is eightfold. It does not mean that the entire database becomes eight times smaller, nor that the saving is free of retrieval-quality trade-offs.

The practical conclusion is to treat the new format as a separate retrieval configuration. Test it on your own collection, reference queries and production-like load. Disk savings matter only when answer quality and latency remain inside an agreed range.

What Qdrant 1.19 adds

Several changes affect both engineering and infrastructure budgets:

  • TurboQuant 4-bit can serve as the primary vector storage type, so a full neighbouring `float32` copy is no longer mandatory;
  • collection components share a common `cold`, `cached` or `pinned` memory strategy;
  • a global quota API is available;
  • sparse retrieval gains request-level IDF statistics, useful in tenant-separated data;
  • batch reads, `io_uring`, `mmap` storage and cgroup resource telemetry were improved.

The release also fixes a path-traversal issue related to S3 snapshots. Teams using that backup mode should therefore assess the upgrade as a security update as well as a storage experiment.

Where the storage saving comes from

A dense embedding with 1,536 dimensions in `float32` occupies 1,536 × 4 bytes, or 6,144 bytes, before the index, identifier and metadata are counted. A four-bit representation needs half a byte per component, so the raw vector occupies roughly 768 bytes.

Consider a simplified model:

  • 10 million vectors;
  • 1,536 dimensions;
  • original `float32` storage;
  • excluding HNSW, payloads, WAL, replicas, service structures and backups.

The raw array is about 61.4 GB in `float32` and 7.7 GB at four bits. The difference is substantial, but the full collection remains larger because graph indexes, filters, text fields, segment versions, replication and snapshots do not disappear.

The useful metric is therefore the cost of a working dataset per million documents at a defined recall, p95 latency and replica count—not the slogan “eight times cheaper.”

Why four bits are not a free speed-up

TurboQuant transforms vectors before quantisation to reduce error. The underlying research describes near-optimal distortion performance without training a codebook, and Qdrant adapts the approach to practical embedding collections.

Compression still introduces a trade-off:

  • less data must be read from disk and more can fit in memory;
  • the transformation and comparison of quantised values require computation;
  • approximate distances may change which documents appear in top-k;
  • results depend on the embedding model and query distribution.

Public benchmarks and theoretical advantages are useful evidence, not substitutes for a test on company documents. A small recall change may be harmless in product recommendations and unacceptable in a policy or compliance assistant.

Where it is most useful

Strong candidates include product catalogues with millions of items and several embeddings, archives of contracts or support tickets whose indexes grow rapidly, multi-tenant RAG with quotas, private environments constrained by NVMe or RAM, and edge or recovery nodes where keeping the full `float32` layer is expensive.

For tens of thousands of passages, the gain may not justify the extra operational choice. Correct chunking, access filtering, hybrid retrieval and document freshness are often more important at that scale.

A safe production test

Do not replace the current collection in place. Run a parallel comparison:

1. Pin the embedding-model version and export a representative dataset.
2. Create a separate four-bit collection without changing production.
3. Build a query set covering common, rare, short, long and domain-specific requests.
4. Measure Recall@k, MRR or nDCG and, for the final RAG response, correct source citation.
5. Compare collection size, working-set RAM, build time, p50/p95 latency and throughput.
6. Test permission filters and multi-tenancy independently from semantic similarity.
7. If top-k quality declines, retrieve more candidates and evaluate a reranker.
8. Restore a snapshot and prove rollback before enabling the format in production.

The reference set must include negative questions, near-duplicate documents, competing document versions and questions for which the corpus has no answer.

How to calculate the business case

Include primary and replica disks, backups and snapshot traffic, RAM or page cache, migration and reindexing time, reranker cost, engineering time and the business cost of an incorrect retrieval result.

The saving becomes tangible when the compact collection permits a smaller server, fewer NVMe drives or more tenants on the same cluster without breaking the SLO. If the corpus is small and an expensive reranker becomes mandatory, the apparent saving may vanish.

Executive takeaway

Qdrant 1.19 turns compression into a storage-architecture choice rather than merely an accelerator placed beside a full vector copy. That is meaningful for growing private RAG deployments.

Decide with four numbers: collection size, Recall@k, p95 latency and restore cost. Start with a parallel collection and a measurable test. The vector layer may enjoy losing weight, but the business still needs a fitting room.