Weaviate 1.27.0
Weaviate 1.27 is here!
Here are the release ⭐️highlights⭐️!
- Big filtered search improvements
- Multi-target vector search enhancement
- Jina V3 embedding support
- Cancel backups
- Other changes
Improved Filtered Search
Weaviate 1.27 introduces a new filtering strategy with big benefits for performance and scalability.
This exciting new search strategy/algorithm comes from our Applied Research team. Weaviate's new filtered search implementation is inspired by the popular ACORN paper, improving on it to make it even better for Weaviate users.
Weaviate's ACORN algorithm will particularly benefit those of you with very large datasets. More specifically, it should significantly improve performance challenges with negatively correlated filtered HNSW searches.
A filtered search combines the structured, rule-based inclusion and exclusion of entities with the vector-based similarity search. The structured part is the filter, and the vector-based part is the search.
In a negatively correlated filtered search, the filter disproportionately removes entities that would be most similar to the search query, as shown below.
This type of filtered search can be slower than other filtered searches. As in the figure, the vector search starts at the "wrong" end of the vector space, where the most similar entities are removed by the filter.
The new ACORN filter strategy in Weaviate 1.27 dramatically improves the filtered vector search performance in situations like this.
How does ACORN work?
The ACORN algorithm improves filtered search performance in three different ways:
- Objects that do not meet the filters are ignored in distance calculations.
- The algorithm reaches the relevant part of the HNSW graph faster, by using a multi-hop approach to evaluate the neighborhood of candidates.
- The entry points are randomly seeded to speed up convergence to the filtered zone.
This implementation is Weaviate's own version of that described in the original paper, with modifications to make it even better for Weaviate users.
How to use ACORN
ACORN can be enabled as the filter strategy by setting filterStrategy for the relevant HNSW vector index in the collection configuration.
At the time of writing, the filterStrategy parameter is not yet supported by all Weaviate clients. It will be added to all clients shortly.
According to our internal tests, the ACORN algorithm generally improves the filtered vector search performances, with the most significant improvements in negatively correlated filtered searches.
Note that there may be a small cost in recall, but our testing has found this to be minimal if not negligible.
We are really excited for you to try it out.
We think that ACORN will benefit most users. But, we especially encourage you to try it out if you have experienced challenges with filtered HNSW searches.
We would love to hear from you about your experiences with ACORN. Please drop by to our forum and tell us about it!
Multi-Target Vector Search Enhancement
In Weaviate 1.27, multi-target vector search has been improved further to allow searching the same vector field multiple times in one query.
These new options provide even more flexibility for comprehensive searches through your data with Weaviate.
Take a look at the multi-target vector search page for more examples.
Weaviate 1.26 introduced multi-target vector searches, where one search query can search across multiple target vectors.
For example, you can search a product object with a query, and compare the query to the product's:
- title vector and
- description vector
In one search!
Jina V3 embedding support
Weaviate's Jina AI embedding integration now supports Jina AI's new V3 embeddings.
How to use the v3 embedding model
From your perspective, all you have to set is the model name, to jina-embeddings-v3 here.
With the v3 model, you can optionally specify the dimensions parameter in the collection configuration with the Jina AI model integration. This will set Weaviate's Jina AI model integration to obtain vectors of the correct length. You can balance your search quality needs here with resource usage.
Note you can combine Weaviate's quantization techniques with this to further manage resource usage.
See the Jina AI model integration page for information on how to use this model.
You can find more information about the model on JinaAI's website.
This model is a multilingual model that remains small and efficient while performing well on standard retrieval benchmarks.
The v3 model adopts binary passage retrieval capabilities, which understands the difference between the database entries (passages) and the user queries.
Weaviate automatically implements this for convenience, embedding objects as passages during import, and as queries during nearText search.
Furthermore, the v3 model produces Matryoshka-style embeddings, This means that the generated embeddings can be truncated to suit the required precision.
Cancel backups
Creating backups is a big part of database administration, especially in a production setting.
In some cases, however, a backup process may be unwanted. Perhaps a backup was triggered by accident, or a backup process may be stuck.
To address this, we have added a backup cancellation feature. Now, an unwanted, ongoing backup process can be stopped instead of consuming valuable CPU cycles.
Backups can be cancelled through a supported Weaviate client, or through the REST endpoint.
Other changes
Case sensitive vectorization
Weaviate integrates with embedding model providers in order to conveniently vectorize data during import.
From v1.27, text is no longer lowercased by default before vectorization.
This change reflects the modern embeddings' capabilities to deal with uppercase (vs lowercase) characters. We expect that any changes to embeddings here will be very small, but we expect this to be closer to how modern embedding models expect inputs to look.
If you prefer the previous behavior where the inputs are lowercased before import, set the LOWERCASE_VECTORIZATION_INPUT environment variable to true.
Google modules renamed
The Weaviate modules that underpin its Google model integrations have been renamed. Historically, they were named after the PaLM generation of models, thus having names text2vec-palm, generative-palm, multi2vec-palm.
They have been renamed to text2vec-google, generative-google, multi2vec-google respectively.
These modules are aliased under-the-hood in Weaviate, so your existing codebase will continue to work. However, we recommend updating your existing code if you are upgrading Weaviate so that your codebase follows the latest conventions.
PQ training set sampling
Product quantization (PQ) is a technique used to compress vectors for efficient storage and retrieval.
PQ requires calculation of centroids using a training set. From 1.27, if you have ingested more objects than the training set size, Weaviate will now sample the training set from the ingested objects. This should make it more likely that the training set is representative of the data.
Reliability and performance improvements
Between 1.26 and 1.27, Weaviate's database and applied research teams have been hard at work, making significant improvements under-the-hood for reliability and performance improvements.
We won't list them all here, but they include:
- improvements to BM25 & hybrid query performance
- data (LSM) store improvements,
- cluster (Raft) improvements,
- tombstone cleanup improvements
and many more.
As always, we recommend updating your version of Weaviate to the latest available patch release (e.g. 1.26.2 ➡️ 1.26.5). And we encourage you to consider upgrading to the latest release (e.g. 1.26.2 ➡️ 1.27.0).
For a full list of changes, see the release notes for each version.
Summary
Enjoy the new features and improvements in Weaviate 1.27. The release is available open-source as always on GitHub, and is available for new Sandboxes on Weaviate Cloud. So please give it a try!
It will be available for Serverless clusters on Weaviate Cloud soon as well.
Thanks for reading, see you next time 👋!