We store not only embeddings but also their attributes, which are more interesting to users since they can use real-world objects. For example, they can identify objects using their real identificators.

Attributes

Data filtering

You can adjust the indexing scheme for specific filter values for a more efficient search.

Sharding

For horizontal scaling of the index during its creation, you can specify the attributes by which the vectors will be distributed into groups. When processing a search query, the results from multiple shards will be automatically combined.

We support changing indexes as new data arrives (delete, change, or add data to the index), in parallel with search queries.

Online changes

Seamless versioning

You can expand multiple indexes over a single data source (vectors and attributes) and seamlessly transition to new versions of indexes. You can expand different indexes with different parameters of the same data.

Dense and Sparse Vector Support

You can work with vectors of any type. For example, to solve word processing problems, you can use bag-of-words and load appropriate sparse vectors into Vektonn.

AkNN support provides you with the best balance between search speed, memory consumption, and search quality.

Approximate kNN support