I migrated SpaceSparseJaccard using this in another branch (using ann-benchmarks with my PR erikbern/ann-benchmarks#239).

The results are quite impressive, before:

After:

Hi @GuilhemN I have reviewed the request. Yes, we can follow this path, but it is a rather substantial change of the C++ API so I need to discuss it with other people. And also if we are doing something like this, don't you think it's easier to have just a float distance instead? An extra template parameter, you know, it would be nice to get rid of dist_t then forever.

Yes, we can follow this path, but it is a rather substantial change of the C++ API so I need to discuss it with other people.

Sure, no problem.

And also if we are doing something like this, don't you think it's easier to have just a float distance instead? An extra template parameter, you know, it would be nice to get rid of dist_t then forever.

Wouldn't that be a bit weird for hamming distance? And also maybe slower than ints?
But I can do it if you want, that would leave us with just one template parameter (that I called dist_uint_t in this proposal, but should probably be renamed to something like data_t).
And it would actually also break compatibility with the former python interface so it would probably require a new major version, right?

Hi @GuilhemN after some consideration, I think I will likely go with your approach. I don't worry about float being inefficient, but using float instead of int will break compatibility. Please, give me some time, I will batch update the library (there are several important PRs) when I have a bit more time.

Thank you!

Hi! Ok thank you, let me know when I should rebase/rework on this PR :)

Hi @GuilhemN I have to postpone things a bit, then I will update the library "en masse". I will let you know if something needs to be changed, but likely it's going to be ok.