Vector-search

Superduper allows users to implement vector-search in their database by either using in-database functionality, or via a sidecar implementation with lance and FastAPI.

Philosophy

In Superduper, from a user point-of-view vector-search isn't a completely different beast than other ways of using the system:

• The vector-preparation is exactly the same as preparing outputs with any model, with the special difference that the outputs are vectors, arrays or tensors.
• Vector-searches are just another type of database query which happen to use the stored vectors.

Algorithm

Here is a schematic of how vector-search works:

Explanation

A vector-search query has the schematic form:

table_or_collection
    .like(Document(<dict-to-search-with>))      # the operand is vectorized using registered models
    .filter_results(*args, **kwargs)            # the results of vector-search are filtered

table_or_collection
    .filter_results(*args, **kwargs)            # the results of vector-search are filtered
    .like(Document(<dict-to-search-with>))      # the operand is vectorized using registered models

...or

The type of such a query is a CompoundSelect. It's 2 parts are the vector-search part (like) and the filtering part (select).

In the first case, the operand of like is dispatched to a model, which converts this into a vector. The vector is compared to previously saved outputs of the same or a paired model (multi-modal). The most similar ids are retrieved. The select part of the query is then transformed to a similar query which searches within the retrieved ids. The full set of results are returned to the client.

Read here about setting up and detailed usage of vector-search.