recall is computed over uniformly random vectors, so the values come out too low to be meaningful (random high-dimensional vectors have no real neighbor structure). We understand from WG discussion that this is a recognized tension (realistic data vs. its CPU/memory cost), and that #489 only added a guard for empty / under-covered ground truth without changing data realism. Filing/commenting mainly to confirm current status and whether any change is planned for v3.0. (If a discussion already tracks this, please point us to it and we'll move the comment there.)
recall is computed over uniformly random vectors, so the values come out too low to be meaningful (random high-dimensional vectors have no real neighbor structure). We understand from WG discussion that this is a recognized tension (realistic data vs. its CPU/memory cost), and that #489 only added a guard for empty / under-covered ground truth without changing data realism. Filing/commenting mainly to confirm current status and whether any change is planned for v3.0. (If a discussion already tracks this, please point us to it and we'll move the comment there.)