Is it possible to do batch search in vss_search or vss_range_search?

[terry-tao-wd]

Hi, we're using sqlite-vss to search embeddings. I'm wondering if we can pass a list of vectors to vss_search (or vss_range_search), and get back the nearest vectors of each of the passed-in vector?

I tried to pass in a list of vectors in json function, like this:

select rowid, distance
from foo
where vss_search(bar, json('[[1, 2, 3], [3, 2, 1]]'))
limit 20;

But got back this error:

Runtime error: 1st argument is not a vector

Is it that I'm doing it incorrectly or sqlite-vss does not support this feature right now? Thanks a lot!

[asg017]

hey, not currently supported. You might be able to do a JOIN and split your input vectors with json_each(), but I haven't tested it.

Faiss has good support for searching multiple vectors at one, I'll see how hard it would be to add. Will track in #109

[terry-tao-wd]

Hi, thanks a lot for the quick response! I just figured out a way to accomplish this, by union-ing the queries, something like:

(select rowid, distance, 1 as 'qid'
from foo
where vss_search(bar, json('[1, 2, 3]'))
limit 20;)
union
(select rowid, distance, 2 as 'qid'
from foo
where vss_search(bar, json('[3, 1, 1]'))
limit 20;)

and then group the results by the extra 'qid' column. Although I'm not sure if there's a limit to the number of queries that I can union, or other limit, have to do some tests.

But it would definitely be good if sqlite-vss supports it natively. Thanks!