← Changelog
FixedApps

Vector indexes build on big models

A vector index on a model with more than a few tens of thousands of vectors failed to build: the build runs for minutes, and the connection it ran on was dropped after one. Now each vector index builds on a connection of its own with more memory, at a few minutes per 100,000 vectors. An index already marked failed is built again within the hour, or, once it has been given up on, when the schema is saved again.

We measured search on 100,000 vectors of 1,536 dimensions. Without an index it takes about a second and a half; with one, an unfiltered search takes a fraction of a second and finds the same records nearly every time. A filter that keeps only a few percent of the records misses some of the nearest through the index, so the docs now suggest indexing a filter's field too when searches always use it.

More in Vector search.