usearch 0.19.3 → 0.21.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (19) hide show
  1. package/.github/workflows/prerelease.yml +27 -0
  2. package/Cargo.toml +1 -1
  3. package/README.md +21 -2
  4. package/c/test.c +238 -126
  5. package/conanfile.py +1 -1
  6. package/cpp/CMakeLists.txt +6 -2
  7. package/cpp/bench.cpp +41 -4
  8. package/cpp/test.cpp +12 -12
  9. package/docs/benchmarks.md +14 -12
  10. package/include/usearch/index.hpp +603 -100
  11. package/include/usearch/index_punned_dense.hpp +68 -10
  12. package/include/usearch/index_punned_helpers.hpp +30 -3
  13. package/package.json +1 -1
  14. package/python/lib.cpp +57 -5
  15. package/python/scripts/join.py +151 -0
  16. package/python/scripts/test.py +19 -0
  17. package/python/usearch/eval.py +11 -3
  18. package/python/usearch/index.py +119 -10
  19. package/python/usearch/join.py +0 -98
package/cpp/test.cpp CHANGED
@@ -9,8 +9,6 @@
9
9
  using namespace unum::usearch;
10
10
  using namespace unum;
11
11
 
12
- using point_id_t = std::int64_t;
13
-
14
12
  void expect(bool must_be_true) {
15
13
  if (!must_be_true)
16
14
  throw std::runtime_error("Failed!");
@@ -22,6 +20,7 @@ template <typename scalar_at, typename index_at> void test3d(index_at&& index) {
22
20
  using view_t = span_gt<scalar_t const>;
23
21
  using index_t = typename std::remove_reference<index_at>::type;
24
22
  using distance_t = typename index_t::distance_t;
23
+ using label_t = typename index_t::label_t;
25
24
 
26
25
  scalar_t vec[3] = {10, 20, 15};
27
26
  scalar_t vec_a[3] = {15, 16, 17};
@@ -31,7 +30,7 @@ template <typename scalar_at, typename index_at> void test3d(index_at&& index) {
31
30
  index.add(42, view_t{&vec[0], 3ul});
32
31
 
33
32
  // Default approximate search
34
- point_id_t matched_labels[10] = {0};
33
+ label_t matched_labels[10] = {0};
35
34
  distance_t matched_distances[10] = {0};
36
35
  std::size_t matched_count = index.search(view_t{&vec[0], 3ul}, 5).dump_to(matched_labels, matched_distances);
37
36
 
@@ -52,7 +51,7 @@ template <typename scalar_at, typename index_at> void test3d(index_at&& index) {
52
51
  // Validate scans
53
52
  std::size_t count = 0;
54
53
  for (auto member : index) {
55
- point_id_t id = member.label;
54
+ label_t id = member.label;
56
55
  expect(id >= 42 && id <= 44);
57
56
  count++;
58
57
  }
@@ -146,11 +145,12 @@ int main(int, char**) {
146
145
  // static_assert(!std::is_same<index_gt<pearson_correlation_gt<>>::value_type, std::true_type>());
147
146
  // static_assert(!std::is_same<index_gt<haversine_gt<>>::value_type, std::true_type>());
148
147
 
149
- test3d<float>(index_gt<cos_gt<float>, point_id_t, std::uint32_t>{});
150
- test3d<float>(index_gt<l2sq_gt<float>, point_id_t, std::uint32_t>{});
148
+ using big_point_id_t = std::int64_t;
149
+ test3d<float>(index_gt<cos_gt<float>, big_point_id_t, std::uint32_t>{});
150
+ test3d<float>(index_gt<l2sq_gt<float>, big_point_id_t, std::uint32_t>{});
151
151
 
152
- test3d<double>(index_gt<cos_gt<double>, point_id_t, std::uint32_t>{});
153
- test3d<double>(index_gt<l2sq_gt<double>, point_id_t, std::uint32_t>{});
152
+ test3d<double>(index_gt<cos_gt<double>, big_point_id_t, std::uint32_t>{});
153
+ test3d<double>(index_gt<l2sq_gt<double>, big_point_id_t, std::uint32_t>{});
154
154
 
155
155
  test3d<float>(punned_small_t::make(3, metric_kind_t::cos_k));
156
156
  test3d<float>(punned_small_t::make(3, metric_kind_t::l2sq_k));
@@ -161,11 +161,11 @@ int main(int, char**) {
161
161
  test3d_punned<float>(punned_small_t::make(3, metric_kind_t::cos_k));
162
162
  test3d_punned<float>(punned_small_t::make(3, metric_kind_t::l2sq_k));
163
163
 
164
- test_sets(index_gt<jaccard_gt<std::int32_t, float>, point_id_t, std::uint32_t>{});
165
- test_sets(index_gt<jaccard_gt<std::int64_t, float>, point_id_t, std::uint32_t>{});
164
+ test_sets(index_gt<jaccard_gt<std::int32_t, float>, big_point_id_t, std::uint32_t>{});
165
+ test_sets(index_gt<jaccard_gt<std::int64_t, float>, big_point_id_t, std::uint32_t>{});
166
166
 
167
- test_sets_moved<index_gt<jaccard_gt<std::int32_t, float>, point_id_t, std::uint32_t>>();
168
- test_sets_moved<index_gt<jaccard_gt<std::int64_t, float>, point_id_t, std::uint32_t>>();
167
+ test_sets_moved<index_gt<jaccard_gt<std::int32_t, float>, big_point_id_t, std::uint32_t>>();
168
+ test_sets_moved<index_gt<jaccard_gt<std::int64_t, float>, big_point_id_t, std::uint32_t>>();
169
169
 
170
170
  return 0;
171
171
  }
package/docs/benchmarks.md CHANGED
@@ -117,21 +117,23 @@ OPTIONS
117
117
  BigANN benchmark is a good starting point, if you are searching for large collections of high-dimensional vectors.
118
118
  Those often come with precomputed ground-truth neighbors, which is handy for recall evaluation.
119
119
 
120
- | Dataset | Scalar Type | Dimensions | Metric | Size |
121
- | :-------------------------------------- | :---------: | :--------: | :----: | :-------: |
122
- | [Unum UForm Wiki][unum-wiki] | float32 | 256 | IP | 1 GB |
123
- | [Yandex Text-to-Image Sample][unum-t2i] | float32 | 200 | Cos | 1 GB |
124
- | | | | | |
125
- | [Microsoft SPACEV][spacev] | int8 | 100 | L2 | 93 GB |
126
- | [Microsoft Turing-ANNS][turing] | float32 | 100 | L2 | 373 GB |
127
- | [Yandex Deep1B][deep] | float32 | 96 | L2 | 358 GB |
128
- | [Yandex Text-to-Image][t2i] | float32 | 200 | Cos | 750 GB |
129
- | | | | | |
130
- | [ViT-L/12 LAION][laion] | float32 | 2048 | Cos | 2 - 10 TB |
120
+ | Dataset | Scalar Type | Dimensions | Metric | Size |
121
+ | :----------------------------------------- | :---------: | :--------: | :----: | :-------: |
122
+ | [Unum UForm Creative Captions][unum-cc-3m] | float32 | 256 | IP | 3 GB |
123
+ | [Unum UForm Wiki][unum-wiki-1m] | float32 | 256 | IP | 1 GB |
124
+ | [Yandex Text-to-Image Sample][unum-t2i] | float32 | 200 | Cos | 1 GB |
125
+ | | | | | |
126
+ | [Microsoft SPACEV][spacev] | int8 | 100 | L2 | 93 GB |
127
+ | [Microsoft Turing-ANNS][turing] | float32 | 100 | L2 | 373 GB |
128
+ | [Yandex Deep1B][deep] | float32 | 96 | L2 | 358 GB |
129
+ | [Yandex Text-to-Image][t2i] | float32 | 200 | Cos | 750 GB |
130
+ | | | | | |
131
+ | [ViT-L/12 LAION][laion] | float32 | 2048 | Cos | 2 - 10 TB |
131
132
 
132
133
  Luckily, smaller samples of those datasets are available.
133
134
 
134
- [unum-wiki]: https://huggingface.co/datasets/unum-cloud/ann-wiki-1m
135
+ [unum-cc-3m]: https://huggingface.co/datasets/unum-cloud/ann-cc-3m
136
+ [unum-wiki-1m]: https://huggingface.co/datasets/unum-cloud/ann-wiki-1m
135
137
  [unum-t2i]: https://huggingface.co/datasets/unum-cloud/ann-t2i-1m
136
138
  [spacev]: https://github.com/microsoft/SPTAG/tree/main/datasets/SPACEV1B
137
139
  [turing]: https://learning2hash.github.io/publications/microsoftturinganns1B/