Weaviate 1.40 Release
By Steven Van ·
Weaviate's open-source and Cloud release makes vector dropping and 4-bit RQ generally available, and flags one breaking change to defaults.
Weaviate 1.40 is out for the open-source server and on Weaviate Cloud. Two features reach general availability, two are new, and one change to vector configuration defaults is flagged as breaking. The details below are from the announcement for 4-bit Rotational Quantization and the 1.40 release note.
Release highlights
- Dropping a named vector is generally available. The drop vector index endpoint has existed since v1.37 as an experimental feature that was off by default, and in v1.40 it is on by default.
- 4-bit Rotational Quantization (RQ) is generally available.
- The HFresh index now accepts multi-vector embeddings through MUVERA.
- PathSeer joins ACORN and sweeping as a third filter strategy for HNSW.
- The release note also covers changes to the Search REST API (Preview) over the 1.39 patch line.
Dropping a named vector
Dropping a vector deletes its index and removes its embedding from every object. The collection's other named vectors and all object data stay as they are. Before this, the only way to reclaim the disk and index memory held by an unused named vector was to re-create the collection and re-import everything.
A drop cannot be undone. To get the vector back, add it again and then re-vectorize or re-import the objects.
With the Python client (4.24.0 or newer), call delete_vector_index with the vector's name. Over REST, send a DELETE request to /v1/schema/Article/vectors/body_vector/index, where Article is the collection and body_vector the vector. The call returns once Weaviate has accepted the drop. Cleanup then runs in the background, shard by shard, and resumes after a restart.
4-bit Rotational Quantization
4-bit RQ was added in Weaviate 1.39 as a bits setting on the existing RQ quantizer, and it is generally available in 1.40. Weaviate reports that it gives similar recall to 8-bit RQ with a 45% heap reduction on a 1536-dimension dataset. The saving is less than half because HNSW graph metadata also uses memory.
The 1.39 work also sped up the rest of the RQ family:
- SIMD support for the Fast Walsh-Hadamard Transform used for rotation, which is 3.1× to 6.1× faster in the vendor's benchmarks.
- 8-bit RQ encoding is 3.8× faster on an Intel Xeon 8581C and 2.4× faster on an Apple M1.
- Prefetching hints add 7 to 11% query throughput and 12% faster imports on a 1M-vector index. A prefetching bug in AMD64 was also fixed.
- Imports on the 1536-dimension dataset take 16% less time for RQ8 than in 1.38. RQ4 and centered RQ4 come in 37% and 32% under the 1.38 baseline.
Centering and outliers
An opt-in centering flag encodes vectors relative to a mean fitted at compression time. On the datasets tested, recall@10 rose by 0.1 to 6.1 percentage points. It is opt-in because some embedding models are regularized to remove the mean. Storing the two largest-magnitude coordinates exactly added a further 0.2 to 1.7 points. The mean is fitted on a sample capped at 10,000 vectors by default.
To enable it, set bits to 4 in the rq block of vectorIndexConfig, with centering set to true if wanted. The Python client takes the same options through Configure.VectorIndex.Quantizer.rq(bits=4, centering=True).
Scale and TurboQuant
In a brute-force test on the Sphere corpus (768 dimensions, dot product), recall stayed flat from 1M to 250M vectors. Weaviate notes that ANN index parameters do degrade with scale, and recommends sharding for large datasets. Against TurboQuant at 4 bits, Weaviate's benchmarks show centered RQ4 ahead on 7 of 8 datasets, including the multi-vector mLateOn models, where stock TurboQuant fell sharply.
Before upgrading
If your server sets DEFAULT_VECTORIZER_MODULE, Weaviate says to read the "Vector configuration defaults" section of the release note first, since it describes a breaking change.