usearch 0.22.2 → 1.0.0

package/cpp/test.cpp

@@ -3,8 +3,11 @@
  */
 #include <algorithm>
 #include <stdexcept>
+#include <unordered_map>
 
-#include <usearch/index_punned_dense.hpp>
+#include <usearch/index.hpp>
+#include <usearch/index_dense.hpp>
+#include <usearch/index_plugins.hpp>
 
 using namespace unum::usearch;
 using namespace unum;
@@ -14,93 +17,197 @@ void expect(bool must_be_true) {
         throw std::runtime_error("Failed!");
 }
 
-template <typename scalar_at, typename index_at> void test3d(index_at&& index) {
+template <bool punned_ak, typename index_at, typename scalar_at, typename... extra_args_at>
+void test_cosine(index_at& index, std::vector<std::vector<scalar_at>> const& vectors, extra_args_at&&... args) {
 
     using scalar_t = scalar_at;
-    using view_t = span_gt<scalar_t const>;
-    using index_t = typename std::remove_reference<index_at>::type;
+    using index_t = index_at;
+    using key_t = typename index_t::key_t;
     using distance_t = typename index_t::distance_t;
-    using label_t = typename index_t::label_t;
 
-    scalar_t vec[3] = {10, 20, 15};
-    scalar_t vec_a[3] = {15, 16, 17};
-    scalar_t vec_b[3] = {16, 17, 18};
+    // Generate some keys starting from end,
+    // for three vectors from the dataset
+    key_t const key_max = default_free_value<key_t>() - 1;
+    key_t const key_first = key_max - 0;
+    key_t const key_second = key_max - 1;
+    key_t const key_third = key_max - 2;
+    scalar_t const* vector_first = vectors[0].data();
+    scalar_t const* vector_second = vectors[1].data();
+    scalar_t const* vector_third = vectors[2].data();
+    std::size_t dimensions = vectors[0].size();
 
     index.reserve(10);
-    index.add(42, view_t{&vec[0], 3ul});
+    index.add(key_first, vector_first, args...);
 
     // Default approximate search
-    label_t matched_labels[10] = {0};
+    key_t matched_labels[10] = {0};
     distance_t matched_distances[10] = {0};
-    std::size_t matched_count = index.search(view_t{&vec[0], 3ul}, 5).dump_to(matched_labels, matched_distances);
+    std::size_t matched_count = index.search(vector_first, 5, args...).dump_to(matched_labels, matched_distances);
 
     expect(matched_count == 1);
-    expect(matched_labels[0] == 42);
+    expect(matched_labels[0] == key_first);
     expect(std::abs(matched_distances[0]) < 0.01);
 
     // Add more entries
-    search_config_t search_config;
+    index_search_config_t search_config;
     search_config.exact = true;
-    index.add(43, view_t{&vec_a[0], 3ul});
-    index.add(44, view_t{&vec_b[0], 3ul});
+    index.add(key_second, vector_second, args...);
+    index.add(key_third, vector_third, args...);
     expect(index.size() == 3);
 
     // Perform exact search
-    matched_count = index.search(view_t{&vec[0], 3ul}, 5, search_config).dump_to(matched_labels, matched_distances);
+    matched_count = index.search(vector_first, 5, args..., search_config).dump_to(matched_labels, matched_distances);
 
     // Validate scans
     std::size_t count = 0;
     for (auto member : index) {
-        label_t id = member.label;
-        expect(id >= 42 && id <= 44);
+        key_t id = member.key;
+        expect(id <= key_first && id >= key_third);
         count++;
     }
     expect((count == 3));
     expect((index.stats(0).nodes == 3));
 
+    // Try removals and replacements
+    if constexpr (punned_ak) {
+        using labeling_result_t = typename index_t::labeling_result_t;
+        labeling_result_t result = index.remove(key_third);
+        expect(bool(result));
+        expect(index.size() == 2);
+        index.add(key_third, vector_third, args...);
+        expect(index.size() == 3);
+    }
+
     // Search again over reconstructed index
     index.save("tmp.usearch");
     index.load("tmp.usearch");
-    matched_count = index.search(view_t{&vec[0], 3ul}, 5).dump_to(matched_labels, matched_distances);
+    matched_count = index.search(vector_first, 5, args...).dump_to(matched_labels, matched_distances);
     expect(matched_count == 3);
-    expect(matched_labels[0] == 42);
+    expect(matched_labels[0] == key_first);
     expect(std::abs(matched_distances[0]) < 0.01);
 
+    if constexpr (punned_ak) {
+        std::vector<scalar_t> vec_recovered_from_load(dimensions);
+        index.get(key_second, vec_recovered_from_load.data());
+        expect(std::equal(vector_second, vector_second + dimensions, vec_recovered_from_load.data()));
+    }
+
+    // Try batch requests
+    executor_default_t executor;
+    index.reserve({vectors.size(), executor.size()});
+    executor.execute_bulk(vectors.size() - 3, [&](std::size_t thread, std::size_t task) {
+        index_update_config_t config;
+        config.thread = thread;
+        index.add(key_max - task - 3, vectors[task + 3].data(), args..., config);
+    });
+
     // Search again over mapped index
-    index_metadata("tmp.usearch");
+    // file_head_result_t head = index_dense_metadata("tmp.usearch");
+    // expect(head.size == 3);
     index.view("tmp.usearch");
-    matched_count = index.search(view_t{&vec[0], 3ul}, 5).dump_to(matched_labels, matched_distances);
+    matched_count = index.search(vector_first, 5, args...).dump_to(matched_labels, matched_distances);
     expect(matched_count == 3);
-    expect(matched_labels[0] == 42);
+    expect(matched_labels[0] == key_first);
     expect(std::abs(matched_distances[0]) < 0.01);
 
+    if constexpr (punned_ak) {
+        std::vector<scalar_t> vec_recovered_from_view(dimensions);
+        index.get(key_second, vec_recovered_from_view.data());
+        expect(std::equal(vector_second, vector_second + dimensions, vec_recovered_from_view.data()));
+
+        auto compaction_result = index.compact();
+        expect(bool(compaction_result));
+    }
+
     expect(index.memory_usage() > 0);
     expect(index.stats().max_edges > 0);
+
+    // Check metadata
+    if constexpr (punned_ak) {
+        index_dense_metadata_result_t meta = index_dense_metadata("tmp.usearch");
+        expect(bool(meta));
+    }
 }
 
-template <typename scalar_at, typename index_at> void test3d_punned(index_at&& index) {
+template <typename scalar_at, typename key_at, typename slot_at> //
+void test_cosine(std::size_t collection_size, std::size_t dimensions) {
 
     using scalar_t = scalar_at;
-    using view_t = span_gt<scalar_t const>;
-    using span_t = span_gt<scalar_at>;
+    using key_t = key_at;
+    using slot_t = slot_at;
 
-    scalar_t vec42[3] = {10, 20, 15};
-    scalar_t vec43[3] = {19, 22, 11};
+    using index_typed_t = index_gt<float, key_t, slot_t>;
+    using member_cref_t = typename index_typed_t::member_cref_t;
+    using member_citerator_t = typename index_typed_t::member_citerator_t;
 
-    index.reserve(10);
-    index.add(42, view_t{&vec42[0], 3ul});
-
-    // Reconstruct
-    scalar_t vec42_reconstructed[3] = {0, 0, 0};
-    index.get(42, span_t{&vec42_reconstructed[0], 3ul});
-    expect(vec42_reconstructed[0] == vec42[0]);
-    expect(vec42_reconstructed[1] == vec42[1]);
-    expect(vec42_reconstructed[2] == vec42[2]);
-
-    index.add(43, view_t{&vec43[0], 3ul});
-    expect(index.size() == 2);
-    index.remove(43);
-    expect(index.size() == 1);
+    std::vector<std::vector<scalar_at>> matrix(collection_size);
+    for (std::vector<scalar_at>& vector : matrix) {
+        vector.resize(dimensions);
+        std::generate(vector.begin(), vector.end(), [=] { return float(std::rand()) / float(INT_MAX); });
+    }
+
+    struct metric_t {
+        std::vector<std::vector<scalar_at>> const* matrix_ptr;
+        std::size_t dimensions;
+
+        scalar_t const* row(std::size_t i) const noexcept { return (*matrix_ptr)[i].data(); }
+
+        float operator()(member_cref_t const& a, member_cref_t const& b) const {
+            return metric_cos_gt<scalar_t>{}(row(get_slot(b)), row(get_slot(a)), dimensions);
+        }
+        float operator()(scalar_t const* some_vector, member_cref_t const& member) const {
+            return metric_cos_gt<scalar_t>{}(some_vector, row(get_slot(member)), dimensions);
+        }
+        float operator()(member_citerator_t const& a, member_citerator_t const& b) const {
+            return metric_cos_gt<scalar_t>{}(row(get_slot(b)), row(get_slot(a)), dimensions);
+        }
+        float operator()(scalar_t const* some_vector, member_citerator_t const& member) const {
+            return metric_cos_gt<scalar_t>{}(some_vector, row(get_slot(member)), dimensions);
+        }
+    };
+
+    // Template:
+    for (std::size_t connectivity : {3, 13, 50}) {
+        std::printf("- templates with connectivity %zu \n", connectivity);
+        metric_t metric{&matrix, dimensions};
+        index_config_t config(connectivity);
+        index_typed_t index_typed(config);
+        test_cosine<false>(index_typed, matrix, metric);
+    }
+
+    // Type-punned:
+    for (std::size_t connectivity : {3, 13, 50}) {
+        std::printf("- punned with connectivity %zu \n", connectivity);
+        using index_t = index_dense_gt<key_t, slot_t>;
+        metric_punned_t metric(dimensions, metric_kind_t::cos_k, scalar_kind<scalar_at>());
+        index_config_t config(connectivity);
+        index_t index = index_t::make(metric, config);
+        test_cosine<true>(index, matrix);
+    }
+}
+
+template <typename key_at, typename slot_at> void test_tanimoto(std::size_t dimensions, std::size_t connectivity) {
+
+    using key_t = key_at;
+    using slot_t = slot_at;
+
+    using index_punned_t = index_dense_gt<key_t, slot_t>;
+    std::size_t words = divide_round_up<CHAR_BIT>(dimensions);
+    metric_punned_t metric(words, metric_kind_t::tanimoto_k, scalar_kind_t::b1x8_k);
+    index_config_t config(connectivity);
+    index_punned_t index = index_punned_t::make(metric, config);
+
+    executor_default_t executor;
+    std::size_t batch_size = 1000;
+    std::vector<b1x8_t> scalars(batch_size * index.scalar_words());
+    std::generate(scalars.begin(), scalars.end(), [] { return static_cast<b1x8_t>(std::rand()); });
+
+    index.reserve({batch_size + index.size(), executor.size()});
+    executor.execute_bulk(batch_size, [&](std::size_t thread, std::size_t task) {
+        index_update_config_t config;
+        config.thread = thread;
+        index.add(task + 25000, scalars.data() + index.scalar_words() * task, config);
+    });
 }
 
 template <typename index_at> void test_sets(index_at&& index) {
@@ -121,18 +228,18 @@ template <typename index_at> void test_sets(index_at&& index) {
     expect(index.size() == 3);
 }
 
-template <typename index_at> void test_sets_moved() {
+template <typename index_at> void test_sets_moved(index_config_t const& config) {
     {
-        index_at index;
+        index_at index{config};
         test_sets(index);
     }
     {
-        index_at index;
+        index_at index{config};
         index.reserve(1);
         test_sets(index_at(std::move(index)));
     }
     {
-        index_at index;
+        index_at index{config};
         index.reserve(1);
         index_at index_moved = std::move(index);
         test_sets(index_moved);
@@ -141,37 +248,17 @@ template <typename index_at> void test_sets_moved() {
 
 int main(int, char**) {
 
-    static_assert(!std::is_same<index_gt<ip_gt<>>::value_type, std::true_type>());
-    static_assert(!std::is_same<index_gt<cos_gt<>>::value_type, std::true_type>());
-    static_assert(!std::is_same<index_gt<l2sq_gt<>>::value_type, std::true_type>());
-    // static_assert(!std::is_same<index_gt<hamming_gt<>>::value_type, std::true_type>());
-    // static_assert(!std::is_same<index_gt<tanimoto_gt<>>::value_type, std::true_type>());
-    // static_assert(!std::is_same<index_gt<sorensen_gt<>>::value_type, std::true_type>());
-    // static_assert(!std::is_same<index_gt<jaccard_gt<>>::value_type, std::true_type>());
-    // static_assert(!std::is_same<index_gt<pearson_correlation_gt<>>::value_type, std::true_type>());
-    // static_assert(!std::is_same<index_gt<haversine_gt<>>::value_type, std::true_type>());
-
-    using big_point_id_t = std::int64_t;
-    test3d<float>(index_gt<cos_gt<float>, big_point_id_t, std::uint32_t>{});
-    test3d<float>(index_gt<l2sq_gt<float>, big_point_id_t, std::uint32_t>{});
-
-    test3d<double>(index_gt<cos_gt<double>, big_point_id_t, std::uint32_t>{});
-    test3d<double>(index_gt<l2sq_gt<double>, big_point_id_t, std::uint32_t>{});
-
-    test3d<float>(punned_small_t::make(3, metric_kind_t::cos_k));
-    test3d<float>(punned_small_t::make(3, metric_kind_t::l2sq_k));
-
-    test3d<double>(punned_small_t::make(3, metric_kind_t::cos_k));
-    test3d<double>(punned_small_t::make(3, metric_kind_t::l2sq_k));
-
-    test3d_punned<float>(punned_small_t::make(3, metric_kind_t::cos_k));
-    test3d_punned<float>(punned_small_t::make(3, metric_kind_t::l2sq_k));
-
-    test_sets(index_gt<jaccard_gt<std::int32_t, float>, big_point_id_t, std::uint32_t>{});
-    test_sets(index_gt<jaccard_gt<std::int64_t, float>, big_point_id_t, std::uint32_t>{});
-
-    test_sets_moved<index_gt<jaccard_gt<std::int32_t, float>, big_point_id_t, std::uint32_t>>();
-    test_sets_moved<index_gt<jaccard_gt<std::int64_t, float>, big_point_id_t, std::uint32_t>>();
+    for (std::size_t collection_size : {10, 500})
+        for (std::size_t dimensions : {97, 256}) {
+            std::printf("Indexing %zu vectors with cos: <float, std::int64_t, std::uint32_t> \n", collection_size);
+            test_cosine<float, std::int64_t, std::uint32_t>(collection_size, dimensions);
+            std::printf("Indexing %zu vectors with cos: <float, std::int64_t, uint40_t> \n", collection_size);
+            test_cosine<float, std::int64_t, uint40_t>(collection_size, dimensions);
+        }
+
+    for (std::size_t connectivity : {3, 13, 50})
+        for (std::size_t dimensions : {97, 256})
+            test_tanimoto<std::int64_t, std::uint32_t>(dimensions, connectivity);
 
     return 0;
 }

package/docs/benchmarks.md

@@ -39,24 +39,24 @@ The main columns are:
 | `f32` x256 |      16      |   64   |   32   |   128'644    |     228'422     |         97.2% |
 | `f32` x256 |      16      |  256   |  128   |    39'981    |     69'065      |         99.2% |
 
-### Different vectors "accuracy"
+### Different vectors "quantization"
 
 | Vectors      | Connectivity | EF @ A | EF @ S | **Add**, QPS | **Search**, QPS | **Recall @1** |
 | :----------- | :----------: | :----: | :----: | :----------: | :-------------: | ------------: |
 | `f32` x256   |      16      |  128   |   64   |    87'995    |     171'856     |         99.1% |
 | `f16` x256   |      16      |  128   |   64   |    87'270    |     153'788     |         98.4% |
 | `f16` x256 ✳️ |      16      |  128   |   64   |    71'454    |     132'673     |         98.4% |
-| `f8` x256    |      16      |  128   |   64   |   115'923    |     274'653     |         98.9% |
+| `i8` x256    |      16      |  128   |   64   |   115'923    |     274'653     |         98.9% |
 
-As seen on the chart, for `f16` accuracy, performance may differ depending on native hardware support for that numeric type.
-Also worth noting, 8-bit quantization results in almost no accuracy loss and may perform better than `f16`.
+As seen on the chart, for `f16` quantization, performance may differ depending on native hardware support for that numeric type.
+Also worth noting, 8-bit quantization results in almost no quantization loss and may perform better than `f16`.
 
 ## Utilities
 
 Within this repository you will find two commonly used utilities:
 
 - `cpp/bench.cpp` the produces the `bench` binary for broad USearch benchmarks.
-- `python/bench.py` for simple benchmarks against FAISS.
+- `python/bench.py` and `python/bench.ipynb` for interactive charts against FAISS.
 
 To achieve best highest results we suggest compiling locally for the target architecture.
 
@@ -112,6 +112,27 @@ OPTIONS
         -h, --help  Print this help information on this tool and exit
 ```
 
+Here is an example of running the C++ benchmark:
+
+```sh
+./build_release/bench \
+    --vectors datasets/wiki_1M/base.1M.fbin \
+    --queries datasets/wiki_1M/query.public.100K.fbin \
+    --neighbors datasets/wiki_1M/groundtruth.public.100K.ibin
+
+./build_release/bench \
+    --vectors datasets/t2i_1B/base.1B.fbin \
+    --queries datasets/t2i_1B/query.public.100K.fbin \
+    --neighbors datasets/t2i_1B/groundtruth.public.100K.ibin \
+    --output datasets/t2i_1B/index.usearch \
+    --cos
+```
+
+
+> Optional parameters include `connectivity`, `expansion_add`, `expansion_search`.
+
+For Python, jut open the Jupyter Notebook and start playing around.
+
 ## Datasets
 
 BigANN benchmark is a good starting point, if you are searching for large collections of high-dimensional vectors.

package/docs/c/reference.rst

@@ -1,7 +1,7 @@
 API Reference
 ===============
 
-===============
-usearch
-===============
+==================
+usearch/usearch.h
+==================
 .. doxygenfile:: usearch.h

package/docs/compilation.md

@@ -1,12 +1,13 @@
 # Compiling USearch
 
+This document describes the workflow for advanced USearch users and enthusiasts willing to contribute to the project.
 Before building the first time, please pull submodules.
 
 ```sh
 git submodule update --init --recursive
 ```
 
-## C++
+## C++11
 
 Linux:
 
@@ -43,29 +44,19 @@ cmake -B ./build_release \
 Linting:
 
 ```sh
-cppcheck --enable=all --suppress=cstyleCast --suppress=unusedFunction \
+cppcheck --enable=all --force --suppress=cstyleCast --suppress=unusedFunction \
     include/usearch/index.hpp \
-    include/index_punned_dense.hpp \
-    include/index_punned_helpers.hpp
+    include/index_dense.hpp \
+    include/index_plugins.hpp
 ```
 
-Benchmarking:
+Testing:
 
 ```sh
-./build_release/bench \
-    --vectors datasets/wiki_1M/base.1M.fbin \
-    --queries datasets/wiki_1M/query.public.100K.fbin \
-    --neighbors datasets/wiki_1M/groundtruth.public.100K.ibin
-
-./build_release/bench \
-    --vectors datasets/t2i_1B/base.1B.fbin \
-    --queries datasets/t2i_1B/query.public.100K.fbin \
-    --neighbors datasets/t2i_1B/groundtruth.public.100K.ibin \
-    --output datasets/t2i_1B/index.usearch \
-    --cos
+cmake -DCMAKE_CXX_COMPILER=gcc-12 -DCMAKE_CXX_COMPILER=g++-12 -DCMAKE_BUILD_TYPE=Debug -B ./build_debug && make -C ./build_debug && ./build_debug/test
 ```
 
-## Python
+## Python 3
 
 Use PyTest to validate the build.
 The `-s` option will disable capturing the logs.
@@ -89,30 +80,26 @@ pip install cibuildwheel
 cibuildwheel --platform linux
 ```
 
-Benchmarking:
+## JavaScript
+
+Node.JS:
 
 ```sh
-pip install faiss-cpu
-python python/scripts/bench.py speed \
-    --vectors datasets/wiki_1M/base.1M.fbin \
-    --queries datasets/wiki_1M/query.public.100K.fbin \
-    --neighbors datasets/wiki_1M/groundtruth.public.100K.ibin
+npm install && node javascript/test.js
+npm publish
 ```
 
-> Optional parameters include `connectivity`, `expansion_add`, `expansion_search`.
-
-Checking the effect of different embedding dimensions on construction speed:
+WebAssembly:
 
 ```sh
-python python/scripts/bench.py dimensions ...
-python python/scripts/bench.py connectivity ...
+emcmake cmake -B ./build -DCMAKE_CXX_FLAGS="${CMAKE_CXX_FLAGS} -s TOTAL_MEMORY=64MB" && emmake make -C ./build
+node ./build/test.js
 ```
 
-## JavaScript
+If you don't yet have `emcmake` installed:
 
 ```sh
-npm install && node javascript/test.js
-npm publish
+git clone https://github.com/emscripten-core/emsdk.git && ./emsdk/emsdk install latest && ./emsdk/emsdk activate latest && source ./emsdk/emsdk_env.sh
 ```
 
 ## Rust
@@ -156,34 +143,51 @@ java -cp . -Djava.library.path="$(pwd)/java/cloud/unum/usearch/" Index.java
 java -cp . -Djava.library.path="$(pwd)/java" cloud.unum.usearch.Index
 ```
 
-## Wolfram
+## Objective-C and Swift
 
 ```sh
-brew install --cask wolfram-engine
+swift build
+swift test -v
 ```
 
-## GoLang
+## C 99
+
+There are a few ways to compile the C 99 USearch SDK.
+Using the Makefile:
 
 ```sh
-cd golang
-go test -v
+make -C ./c build
 ```
 
-## Objective-C and Swift
+Using CMake:
 
 ```sh
-swift build
-swift test -v
+cmake -B ./build_release -DUSEARCH_BUILD_C=1 && make -C ./build_release -j
 ```
 
-## C
+Linux:
+
+```sh
+g++ -std=c++11 -shared -fPIC c/lib.cpp -I ./include/  -I ./fp16/include/ -I ./robin-map/include/ -o libusearch.so
+```
+
+
+## GoLang
+
+GoLang bindings are based on C.
+So one should first compile the C library, link it with GoLang, and only then run tests.
 
-### Linux
+```sh
+cd golang && make -C ../c build && mv ../c/libusearch.so libusearch.so && go test -v
+```
+
+## Wolfram
 
 ```sh
-g++ -shared -fPIC lib.cpp -I ../include/  -I ../fp16/include/ -o libusearch.so
+brew install --cask wolfram-engine
 ```
 
+
 ## Docker
 
 ```sh
@@ -205,10 +209,10 @@ docker buildx create --use &&
         --push .
 ```
 
----
+## Sub-Modules
 
 Extending metrics in SimSIMD:
 
-```
+```sh
 git push --set-upstream https://github.com/ashvardanian/simsimd.git HEAD:main
-```
+```

package/docs/cpp/index.rst

@@ -1,6 +1,6 @@
-=======
+=====================
 C++ SDK
-=======
+=====================
 
 
 .. mdinclude:: ../../cpp/README.md

package/docs/cpp/reference.rst

@@ -1,7 +1,17 @@
 API Reference
 ===============
 
-===============
-usearch
-===============
+=================================
+usearch/index.hpp
+=================================
 .. doxygenfile:: index.hpp
+
+=================================
+usearch/index_plugins.hpp
+=================================
+.. doxygenfile:: index_plugins.hpp
+
+=================================
+usearch/index_dense.hpp
+=================================
+.. doxygenfile:: index_dense.hpp

package/docs/index.rst

@@ -3,27 +3,27 @@ Overview
 ==========
 .. mdinclude:: ../README.md
 
-.. toctree:: 
-   :hidden:
-   :caption: �
-
-   benchmarks
-   compilation
-
 .. toctree:: 
    :hidden:
    :caption: �
 
    cpp/index
    python/index
+   rust/index
    javascript/index
    java/index
-   rust/index
    c/index
    swift/index
    golang/index
    wolfram/index
 
+.. toctree:: 
+   :hidden:
+   :caption: �
+
+   compilation
+   benchmarks
+
 .. toctree:: 
    :hidden:
    :caption: �

package/docs/java/reference.rst

@@ -2,6 +2,6 @@ API Reference
 ===============
 
 ===============
-usearch
+Index.java
 ===============
 .. doxygenfile:: Index.java