Is the regression for building or for querying? We should ideally benchmark both. I think we'd be willing to take a small regression on building but not regressing querying is actually more important. Also would be good to take a quick look at the impact on in-memory size.

I've just run some benchmarks for querying. There's a lot of variance in the results, I think it's coming from Python's overhead since I'm running search 10k times and it takes a few seconds. But even with that variance I can clearly see that making the distance calculation not bound to vector length regresses querying slightly and then actually using Vec in vectors regresses it further.

For original the shortest loop took 3.227s, after updating the metric it grew to 5.059s and after using AVec it went up to 5.596s. All those are the shortest times out of 40 runs (each run doing 10k searches).

Right, those are definitely some substantial latency regressions -- so I don't think we want to go this route.

(Probably not necessary right now, but it might be interesting to test the low-level metric stuff with Rust benchmarks instead of the Python bindings? Would get rid of the Python noise at least.)

Should keep this issue in mind:

#35

I think I'd prefer to continue with this one and close/rework #35 once this PR is merged.

Why is that? Do you think you'll still have time to work on this? Hope the new job is working out well for you!

The new job is great. I'm still hoping to finish this but the last few months are proving me wrong.

Ambient looks cool, BTW. How about we just say that we're going to pick it up, and if you get to it before us that's nice, but we'll both assume that you have enough other things on your plate as it is? 🙂