Something about a vector binary embedding of an entire relays content to perhaps make searching a compressed relays content extremely efficient and perhaps you could embed all of the content across the entire protocol. Then it could be redundant and spread everywhere and all nostr content may no longer need to be fragmented across relays but aggregated with the embedding and globally distributed across all relays. Sounds really complex though.
Discussion
I'm hesitant about doing it for kind 1 because there is such a wide context, lots of notes collected over lots of relays - don't think it would really work because for the user it there is so many unrelated notes. But imagine a pool of relays focused on some content:
Even if you group across multiple relays, a user is choosing to connect to related content. A user is reading some related articles that have been embedded and so you (service provider) or the client is able to construct some sort of HNSW search, and pass some sort of subnetwork within some distance close to your note.
I suppose it would depend on how good binary vector embedding actually is. 😂
Its a compressed representation, so you lose some resolution to search through, and the space of topics that can be covered in a feed of kind1 notes over some length of time is huge. Less so with a themed community. So by going with something focused you and only improve in performance.
* you can only improve
Dropping a demo in case you'd want to play around with it
Good point