Secondary indices for vector search

[ml-evs]

In #993, I introduce a search index that has to be manually constructed on each edit of a sample. This is optional, and not particularly resource intensive (though should be benchmarked). The same approach could be taken to enable vector search. As this would be quite niche, I think simply using something like https://github.com/asg017/sqlite-vec could work well, and would allow for much more powerful AI agent use.

cc @jdbocarsly

[ml-evs]

This is also something that could be developed as an "offline" search index in the yellowhammer repo, i.e., download all public data from a deployment, create the index/embedding, then add an sqlite layer on the client-side.