Performance

LanceDB designed our architecture in a way to deliver 25ms vector search latency. Even with metadata filtering, our query latency remains as low as 50ms. It is important to note that we can support thousands of QPS with such query performance.

​

Dataset

We used two datasets for this benchmark test. The dbpedia-entities-openai-1M for vector search, and a synthetic dataset for vector search with metadata filtering.

​

Vector search

We ran vector queries with dbpedia-entities-openai-1M with warmed up cache. The query latency is as follows:

​

Full-text search

With the same dataset, and warmed-up cache, the full-text search performance is as follows:

​

Vector search with metadata filtering

We created a 15M-vector dataset with sufficient complexity to thoroughly test our complex metadata filtering capabilities. Such filtering can span a wide range on the scalar columns, e.g. find Sci-fi movies since 1900.

With a warmed-up cache, the query performance using slightly more selective filters, e.g. find Sci-fi movies between the year of 2000 and 2012, is as follows:

The query performance using complex filters, e.g. find Sci-fi movies since 1900, is as follows: