FR: support vector search

[vsekhar]

Some libsql forks/extensions[1] support vector storage (e.g. data type F32_BLOB(n)), vector indexing (e.g. libsql_vector_idx(v, 'metric=cosine')), and vector functions (e.g. vec_distance_l2(a, b).

Under the hood, DiskANN seems to be the preferred pairing with SQLite (e.g. Turso's implementation).

[1]: Turso and SQLite-vec

[mbrandonw]

Hi @vsekhar, as long as the sqlite-vec extension is available to the version of SQLite you are using, you should already have access to everything you need. We even have a very simple example of what it would look like to write a query that uses sample_embedding match here. Note that this is not actually executing the query, but just shows how to use the #sql macro to get access to any of sqlite-vec's functionality that we do not officially support in the query builder. And of course you can always define your own helpers to add more safety too, just as we did with FTS5 here.

Now, if you need to use a custom SQLite to get the functionality you want, that will be more work. Under the hood we use GRDB for access to SQLite, and that library has this documentation on using custom SQLite's.

If you do try using our libraries with sqlite-vec (or Turso) and you run into any concrete problems, please do let us know. We'd love to take a deeper look.