Totally. I get a lot of that.
This is just me poking at the code, my use cases, and reporting back my findings in a… Rod-ish fashion.
Aside from raw speed comparisons (which will obviously improve from alpha code), most of my nudging in this thread (and general line of inquiry) has to do with…
because
This may change with some things as you’re suggesting, but if any variability of distance matching, biasing, forking, etc… requires copying and re-fit-ing a tree, I think there may be a big gap in terms of what is possible in a real-time context.
I’m imagining (hoping) that for vanilla distance matching, that a pre-fit kdtree would be fast as shit, and be a super go-to where you only want to find the nearest in a set, with fixed data all around (like for the time-travel/prediction stuff). I could be unique in that sense, but for me this covers only a small part of what I’d want to do with a corpus. So a lot of this is just gesturing towards those possibilities, which circles back to…