usearch 0.1.10 → 0.2.2

package/python/lib.cpp

@@ -24,11 +24,41 @@ using namespace unum::usearch;
 using namespace unum;
 
 namespace py = pybind11;
+using py_shape_t = py::array::ShapeContainer;
 
 using label_t = Py_ssize_t;
 using distance_t = punned_distance_t;
+using id_t = std::uint32_t;
 using native_index_t = auto_index_gt<label_t>;
 
+using set_member_t = std::uint32_t;
+using set_view_t = span_gt<set_member_t const>;
+using sets_index_t = index_gt<jaccard_gt<set_member_t>, label_t, id_t, set_member_t>;
+
+using hash_word_t = std::uint64_t;
+using hash_index_t = index_gt<bit_hamming_gt<hash_word_t>, label_t, id_t, hash_word_t>;
+static constexpr std::size_t bits_per_hash_word_k = sizeof(hash_word_t) * CHAR_BIT;
+
+struct hash_index_w_meta_t : public hash_index_t {
+    using hash_index_t::add;
+    using hash_index_t::capacity;
+    using hash_index_t::reserve;
+    using hash_index_t::search;
+    using hash_index_t::size;
+
+    std::vector<hash_word_t> buffer_;
+    std::size_t words_;
+    std::size_t bits_;
+
+    hash_index_w_meta_t(config_t config, std::size_t bits) : hash_index_t(config) {
+        words_ = divide_round_up<bits_per_hash_word_k>(bits);
+        bits_ = words_ * bits_per_hash_word_k;
+        buffer_.resize(words_);
+    }
+
+    span_gt<hash_word_t const> buffer() const noexcept { return {buffer_.data(), buffer_.size()}; }
+};
+
 static native_index_t make_index(   //
     std::size_t dimensions,         //
     std::size_t capacity,           //
@@ -56,33 +86,91 @@ static native_index_t make_index(   //
         return index_from_name<native_index_t>(metric.c_str(), metric.size(), dimensions, accuracy, config);
 }
 
-static void add_to_index(native_index_t& index, py::buffer labels, py::buffer vectors, bool copy) {
+static std::unique_ptr<sets_index_t> make_sets_index( //
+    std::size_t capacity,                             //
+    std::size_t connectivity,                         //
+    std::size_t expansion_add,                        //
+    std::size_t expansion_search                      //
+) {
+    config_t config;
+    config.expansion_add = expansion_add;
+    config.expansion_search = expansion_search;
+    config.connectivity = connectivity;
+    config.max_elements = capacity;
+    config.max_threads_add = 1;
+    config.max_threads_search = 1;
+
+    return std::unique_ptr<sets_index_t>(new sets_index_t(config));
+}
+
+static std::unique_ptr<hash_index_w_meta_t> make_hash_index( //
+    std::size_t bits,                                        //
+    std::size_t capacity,                                    //
+    std::size_t connectivity,                                //
+    std::size_t expansion_add,                               //
+    std::size_t expansion_search                             //
+) {
+    config_t config;
+    config.expansion_add = expansion_add;
+    config.expansion_search = expansion_search;
+    config.connectivity = connectivity;
+    config.max_elements = capacity;
+    config.max_threads_add = 1;
+    config.max_threads_search = 1;
+
+    return std::unique_ptr<hash_index_w_meta_t>(new hash_index_w_meta_t(config, bits));
+}
+
+static void add_one_to_index(native_index_t& index, label_t label, py::buffer vector, bool copy) {
+
+    py::buffer_info vector_info = vector.request();
+    if (vector_info.ndim != 1)
+        throw std::invalid_argument("Expects a vector, not a higher-rank tensor!");
+
+    ssize_t vector_dimensions = vector_info.shape[0];
+    char const* vector_data = reinterpret_cast<char const*>(vector_info.ptr);
+    if (vector_dimensions != static_cast<ssize_t>(index.dimensions()))
+        throw std::invalid_argument("The number of vector dimensions doesn't match!");
+
+    if (index.size() + 1 >= index.capacity())
+        index.reserve(ceil2(index.size() + 1));
+
+    // https://docs.python.org/3/library/struct.html#format-characters
+    if (vector_info.format == "e")
+        index.add(label, reinterpret_cast<f16_bits_t const*>(vector_data), 0, copy);
+    else if (vector_info.format == "f")
+        index.add(label, reinterpret_cast<float const*>(vector_data), 0, copy);
+    else if (vector_info.format == "d")
+        index.add(label, reinterpret_cast<double const*>(vector_data), 0, copy);
+    else
+        throw std::invalid_argument("Incompatible scalars in the vector!");
+}
+
+static void add_many_to_index(native_index_t& index, py::buffer labels, py::buffer vectors, bool copy) {
 
     py::buffer_info labels_info = labels.request();
     py::buffer_info vectors_info = vectors.request();
 
     if (labels_info.format != py::format_descriptor<label_t>::format())
-        throw std::runtime_error("Incompatible label type!");
+        throw std::invalid_argument("Incompatible label type!");
 
     if (labels_info.ndim != 1)
-        throw std::runtime_error("Labels must be placed in a single-dimensional array!");
+        throw std::invalid_argument("Labels must be placed in a single-dimensional array!");
 
     if (vectors_info.ndim != 2)
-        throw std::runtime_error("Expects a matrix of vectors to add!");
+        throw std::invalid_argument("Expects a matrix of vectors to add!");
 
     ssize_t labels_count = labels_info.shape[0];
     ssize_t vectors_count = vectors_info.shape[0];
     ssize_t vectors_dimensions = vectors_info.shape[1];
-    if (vectors_dimensions != index.dimensions())
-        throw std::runtime_error("The number of vector dimensions doesn't match!");
+    if (vectors_dimensions != static_cast<ssize_t>(index.dimensions()))
+        throw std::invalid_argument("The number of vector dimensions doesn't match!");
 
     if (labels_count != vectors_count)
-        throw std::runtime_error("Number of labels and vectors must match!");
+        throw std::invalid_argument("Number of labels and vectors must match!");
 
-    if (index.size() + vectors_count >= index.capacity()) {
-        std::size_t next_capacity = ceil2(index.size() + vectors_count);
-        index.reserve(next_capacity);
-    }
+    if (index.size() + vectors_count >= index.capacity())
+        index.reserve(ceil2(index.size() + vectors_count));
 
     char const* vectors_data = reinterpret_cast<char const*>(vectors_info.ptr);
     char const* labels_data = reinterpret_cast<char const*>(labels_info.ptr);
@@ -91,8 +179,8 @@ static void add_to_index(native_index_t& index, py::buffer labels, py::buffer ve
     if (vectors_info.format == "e")
         multithreaded(index.concurrency(), vectors_count, [&](std::size_t thread_idx, std::size_t task_idx) {
             label_t label = *reinterpret_cast<label_t const*>(labels_data + task_idx * labels_info.strides[0]);
-            f16_converted_t const* vector =
-                reinterpret_cast<f16_converted_t const*>(vectors_data + task_idx * vectors_info.strides[0]);
+            f16_bits_t const* vector =
+                reinterpret_cast<f16_bits_t const*>(vectors_data + task_idx * vectors_info.strides[0]);
             index.add(label, vector, thread_idx, copy);
         });
     else if (vectors_info.format == "f")
@@ -108,11 +196,48 @@ static void add_to_index(native_index_t& index, py::buffer labels, py::buffer ve
             index.add(label, vector, thread_idx, copy);
         });
     else
-        throw std::runtime_error("Incompatible scalars in the vectors matrix!");
+        throw std::invalid_argument("Incompatible scalars in the vectors matrix!");
+}
+
+static py::tuple search_one_in_index(native_index_t& index, py::buffer vector, std::size_t wanted) {
+
+    py::buffer_info vector_info = vector.request();
+    ssize_t vector_dimensions = vector_info.shape[0];
+    char const* vector_data = reinterpret_cast<char const*>(vector_info.ptr);
+    if (vector_dimensions != static_cast<ssize_t>(index.dimensions()))
+        throw std::invalid_argument("The number of vector dimensions doesn't match!");
+
+    py::array_t<label_t> labels_py(static_cast<ssize_t>(wanted));
+    py::array_t<distance_t> distances_py(static_cast<ssize_t>(wanted));
+    std::size_t count{};
+    auto labels_py1d = labels_py.mutable_unchecked<1>();
+    auto distances_py1d = distances_py.mutable_unchecked<1>();
+
+    // https://docs.python.org/3/library/struct.html#format-characters
+    if (vector_info.format == "e")
+        count = index.search( //
+            reinterpret_cast<f16_bits_t const*>(vector_data), wanted, &labels_py1d(0), &distances_py1d(0), 0);
+    else if (vector_info.format == "f")
+        count = index.search( //
+            reinterpret_cast<float const*>(vector_data), wanted, &labels_py1d(0), &distances_py1d(0), 0);
+    else if (vector_info.format == "d")
+        count = index.search( //
+            reinterpret_cast<double const*>(vector_data), wanted, &labels_py1d(0), &distances_py1d(0), 0);
+    else
+        throw std::invalid_argument("Incompatible scalars in the query vector!");
+
+    labels_py.resize(py_shape_t{static_cast<ssize_t>(count)});
+    distances_py.resize(py_shape_t{static_cast<ssize_t>(count)});
+
+    py::tuple results(3);
+    results[0] = labels_py;
+    results[1] = distances_py;
+    results[2] = static_cast<Py_ssize_t>(count);
+    return results;
 }
 
 /**
- *  @param vectors Matrix of vectors to search for..
+ *  @param vectors Matrix of vectors to search for.
  *  @param wanted Number of matches per request.
  *
  *  @return Tuple with:
@@ -120,30 +245,34 @@ static void add_to_index(native_index_t& index, py::buffer labels, py::buffer ve
  *      2. matrix of distances,
  *      3. array with match counts.
  */
-static py::tuple search_in_index(native_index_t& index, py::buffer vectors, ssize_t wanted) {
+static py::tuple search_many_in_index(native_index_t& index, py::buffer vectors, std::size_t wanted) {
+
+    if (wanted == 0)
+        return py::tuple(3);
 
     py::buffer_info vectors_info = vectors.request();
+    if (vectors_info.ndim == 1)
+        return search_one_in_index(index, vectors, wanted);
     if (vectors_info.ndim != 2)
-        throw std::runtime_error("Expects a matrix of vectors to add!");
+        throw std::invalid_argument("Expects a matrix of vectors to add!");
 
     ssize_t vectors_count = vectors_info.shape[0];
     ssize_t vectors_dimensions = vectors_info.shape[1];
-    if (vectors_dimensions != index.dimensions())
-        throw std::runtime_error("The number of vector dimensions doesn't match!");
+    char const* vectors_data = reinterpret_cast<char const*>(vectors_info.ptr);
+    if (vectors_dimensions != static_cast<ssize_t>(index.dimensions()))
+        throw std::invalid_argument("The number of vector dimensions doesn't match!");
 
-    py::array_t<label_t> labels_py({vectors_count, wanted});
-    py::array_t<distance_t> distances_py({vectors_count, wanted});
+    py::array_t<label_t> labels_py({vectors_count, static_cast<ssize_t>(wanted)});
+    py::array_t<distance_t> distances_py({vectors_count, static_cast<ssize_t>(wanted)});
     py::array_t<Py_ssize_t> counts_py(vectors_count);
     auto labels_py2d = labels_py.mutable_unchecked<2>();
     auto distances_py2d = distances_py.mutable_unchecked<2>();
     auto counts_py1d = counts_py.mutable_unchecked<1>();
 
-    char const* vectors_data = reinterpret_cast<char const*>(vectors_info.ptr);
-
     // https://docs.python.org/3/library/struct.html#format-characters
     if (vectors_info.format == "e")
         multithreaded(index.concurrency(), vectors_count, [&](std::size_t thread_idx, std::size_t task_idx) {
-            f16_converted_t const* vector = (f16_converted_t const*)(vectors_data + task_idx * vectors_info.strides[0]);
+            f16_bits_t const* vector = (f16_bits_t const*)(vectors_data + task_idx * vectors_info.strides[0]);
             counts_py1d(task_idx) = static_cast<Py_ssize_t>(
                 index.search(vector, wanted, &labels_py2d(task_idx, 0), &distances_py2d(task_idx, 0), thread_idx));
         });
@@ -160,7 +289,7 @@ static py::tuple search_in_index(native_index_t& index, py::buffer vectors, ssiz
                 index.search(vector, wanted, &labels_py2d(task_idx, 0), &distances_py2d(task_idx, 0), thread_idx));
         });
     else
-        throw std::runtime_error("Incompatible scalars in the query matrix!");
+        throw std::invalid_argument("Incompatible scalars in the query matrix!");
 
     py::tuple results(3);
     results[0] = labels_py;
@@ -169,18 +298,87 @@ static py::tuple search_in_index(native_index_t& index, py::buffer vectors, ssiz
     return results;
 }
 
-static void save_index(native_index_t const& index, std::string const& path) { index.save(path.c_str()); }
-static void load_index(native_index_t& index, std::string const& path) { index.load(path.c_str()); }
-static void view_index(native_index_t& index, std::string const& path) { index.view(path.c_str()); }
+// clang-format off
+template <typename index_at> void save_index(index_at const& index, std::string const& path) { index.save(path.c_str()); }
+template <typename index_at> void load_index(index_at& index, std::string const& path) { index.load(path.c_str()); }
+template <typename index_at> void view_index(index_at& index, std::string const& path) { index.view(path.c_str()); }
+template <typename index_at> void clear_index(index_at& index) { index.clear(); }
+// clang-format on
+
+template <typename element_at> bool has_duplicates(element_at const* begin, element_at const* end) {
+    if (begin == end)
+        return false;
+    element_at const* last = begin;
+    begin++;
+    for (; begin != end; ++begin, ++last) {
+        if (*begin == *last)
+            return true;
+    }
+    return false;
+}
+
+void validate_set(py::array_t<set_member_t> const& set) {
+    if (set.ndim() != 1)
+        throw std::invalid_argument("Set can't be multi-dimensional!");
+    if (set.strides(0) != sizeof(set_member_t))
+        throw std::invalid_argument("Set can't be strided!");
+    auto proxy = set.unchecked<1>();
+    set_member_t const* begin = &proxy(0);
+    set_member_t const* end = begin + proxy.size();
+    if (!std::is_sorted(begin, end))
+        throw std::invalid_argument("Set must be sorted!");
+    if (has_duplicates(begin, end))
+        throw std::invalid_argument("Set must be deduplicated!");
+}
 
-PYBIND11_MODULE(usearch, m) {
+inline std::uint64_t hash_ror64(std::uint64_t v, int r) noexcept { return (v >> r) | (v << (64 - r)); }
+
+inline std::uint64_t hash(std::uint64_t v) noexcept {
+    v ^= hash_ror64(v, 25) ^ hash_ror64(v, 50);
+    v *= 0xA24BAED4963EE407UL;
+    v ^= hash_ror64(v, 24) ^ hash_ror64(v, 49);
+    v *= 0x9FB21C651E98DF25UL;
+    return v ^ v >> 28;
+}
+
+template <typename scalar_at>
+inline void hash_typed_buffer(hash_index_w_meta_t& index, py::buffer_info const& vector_info) noexcept {
+    char const* vector_data = reinterpret_cast<char const*>(vector_info.ptr);
+    ssize_t vector_dimensions = vector_info.shape[0];
+    ssize_t vector_stride = vector_info.strides[0];
+    std::memset(index.buffer_.data(), 0, index.words_ * sizeof(hash_word_t));
+
+    for (std::size_t i = 0; i != static_cast<std::size_t>(vector_dimensions); ++i) {
+        scalar_at scalar = *reinterpret_cast<scalar_at const*>(vector_data + i * vector_stride);
+        std::uint64_t scalar_hash = hash(scalar);
+        index.buffer_[scalar_hash % index.words_] |= hash_word_t(1) << (scalar_hash % bits_per_hash_word_k);
+    }
+}
+
+void hash_buffer(hash_index_w_meta_t& index, py::buffer vector) {
+    py::buffer_info info = vector.request();
+    if (info.ndim != 1)
+        throw std::invalid_argument("Array can't be multi-dimensional!");
+
+    // https://docs.python.org/3/library/struct.html#format-characters
+    if (info.format == "h" || info.format == "H")
+        return hash_typed_buffer<std::uint16_t>(index, info);
+    else if (info.format == "i" || info.format == "I" || info.format == "l" || info.format == "L")
+        return hash_typed_buffer<std::uint32_t>(index, info);
+    else if (info.format == "q" || info.format == "Q" || info.format == "n" || info.format == "N")
+        return hash_typed_buffer<std::uint64_t>(index, info);
+    else
+        throw std::invalid_argument("Array elements must be 16, 32, or 64 bit hashable integers!");
+}
+
+PYBIND11_MODULE(index, m) {
     m.doc() = "Unum USearch Python bindings";
 
     auto i = py::class_<native_index_t>(m, "Index");
 
     i.def(py::init(&make_index),                                              //
           py::kw_only(),                                                      //
-          py::arg("ndim"),                                                    //
+          py::arg("ndim") = 0,                                                //
           py::arg("capacity") = 0,                                            //
           py::arg("dtype") = std::string("f32"),                              //
           py::arg("metric") = std::string("ip"),                              //
@@ -190,18 +388,26 @@ PYBIND11_MODULE(usearch, m) {
           py::arg("metric_pointer") = 0                                       //
     );
 
-    i.def(                     //
-        "add", &add_to_index,  //
-        py::arg("labels"),     //
-        py::arg("vectors"),    //
-        py::kw_only(),         //
-        py::arg("copy") = true //
+    i.def(                         //
+        "add", &add_many_to_index, //
+        py::arg("labels"),         //
+        py::arg("vectors"),        //
+        py::kw_only(),             //
+        py::arg("copy") = true     //
+    );
+
+    i.def(                        //
+        "add", &add_one_to_index, //
+        py::arg("label"),         //
+        py::arg("vector"),        //
+        py::kw_only(),            //
+        py::arg("copy") = true    //
     );
 
-    i.def(                          //
-        "search", &search_in_index, //
-        py::arg("vectors"),         //
-        py::arg("count") = 10       //
+    i.def(                               //
+        "search", &search_many_in_index, //
+        py::arg("query"),                //
+        py::arg("count") = 10            //
     );
 
     i.def("__len__", &native_index_t::size);
@@ -210,7 +416,111 @@ PYBIND11_MODULE(usearch, m) {
     i.def_property_readonly("connectivity", &native_index_t::connectivity);
     i.def_property_readonly("capacity", &native_index_t::capacity);
 
-    i.def("save", &save_index, py::arg("path"));
-    i.def("load", &load_index, py::arg("path"));
-    i.def("view", &view_index, py::arg("path"));
+    i.def("save", &save_index<native_index_t>, py::arg("path"));
+    i.def("load", &load_index<native_index_t>, py::arg("path"));
+    i.def("view", &view_index<native_index_t>, py::arg("path"));
+    i.def("clear", &clear_index<native_index_t>);
+
+    auto si = py::class_<sets_index_t>(m, "SetsIndex");
+
+    si.def(                                                                //
+        py::init(&make_sets_index),                                        //
+        py::kw_only(),                                                     //
+        py::arg("capacity") = 0,                                           //
+        py::arg("connectivity") = config_t::connectivity_default_k,        //
+        py::arg("expansion_add") = config_t::expansion_add_default_k,      //
+        py::arg("expansion_search") = config_t::expansion_search_default_k //
+    );
+
+    si.def( //
+        "add",
+        [](sets_index_t& index, label_t label, py::array_t<set_member_t> set, bool copy) {
+            validate_set(set);
+            if (index.size() + 1 >= index.capacity())
+                index.reserve(ceil2(index.size() + 1));
+            auto proxy = set.unchecked<1>();
+            auto view = set_view_t{proxy.data(0), static_cast<std::size_t>(proxy.shape(0))};
+            index.add(label, view, 0, copy);
+        },                     //
+        py::arg("label"),      //
+        py::arg("set"),        //
+        py::kw_only(),         //
+        py::arg("copy") = true //
+    );
+
+    si.def( //
+        "search",
+        [](sets_index_t& index, py::array_t<set_member_t> set, std::size_t count) -> py::array_t<label_t> {
+            validate_set(set);
+            auto proxy = set.unchecked<1>();
+            auto view = set_view_t{proxy.data(0), static_cast<std::size_t>(proxy.shape(0))};
+            auto labels_py = py::array_t<label_t>(py_shape_t{static_cast<ssize_t>(count)});
+            auto labels_proxy = labels_py.mutable_unchecked<1>();
+            auto found = index.search(view, count, &labels_proxy(0), nullptr, 0);
+            labels_py.resize(py_shape_t{static_cast<ssize_t>(found)});
+            return labels_py;
+        },
+        py::arg("set"),       //
+        py::arg("count") = 10 //
+    );
+
+    si.def("__len__", &sets_index_t::size);
+    si.def_property_readonly("size", &sets_index_t::size);
+    si.def_property_readonly("connectivity", &sets_index_t::connectivity);
+    si.def_property_readonly("capacity", &sets_index_t::capacity);
+
+    si.def("save", &save_index<sets_index_t>, py::arg("path"));
+    si.def("load", &load_index<sets_index_t>, py::arg("path"));
+    si.def("view", &view_index<sets_index_t>, py::arg("path"));
+    si.def("clear", &clear_index<sets_index_t>);
+
+    auto hi = py::class_<hash_index_w_meta_t>(m, "HashIndex");
+
+    hi.def(                                                                //
+        py::init(&make_hash_index),                                        //
+        py::kw_only(),                                                     //
+        py::arg("bits"),                                                   //
+        py::arg("capacity") = 0,                                           //
+        py::arg("connectivity") = config_t::connectivity_default_k,        //
+        py::arg("expansion_add") = config_t::expansion_add_default_k,      //
+        py::arg("expansion_search") = config_t::expansion_search_default_k //
+    );
+
+    hi.def( //
+        "add",
+        [](hash_index_w_meta_t& index, label_t label, py::buffer array) {
+            if (index.size() + 1 >= index.capacity())
+                index.reserve(ceil2(index.size() + 1));
+            hash_buffer(index, array);
+            index.add(label, index.buffer(), 0);
+        },                //
+        py::arg("label"), //
+        py::arg("array")  //
+    );
+
+    hi.def( //
+        "search",
+        [](hash_index_w_meta_t& index, py::buffer array, std::size_t count) -> py::array_t<label_t> {
+            if (index.size() + 1 >= index.capacity())
+                index.reserve(ceil2(index.size() + 1));
+            hash_buffer(index, array);
+            auto labels_py = py::array_t<label_t>(py_shape_t{static_cast<ssize_t>(count)});
+            auto labels_proxy = labels_py.mutable_unchecked<1>();
+            auto found = index.search(index.buffer(), count, &labels_proxy(0), nullptr, 0);
+            labels_py.resize(py_shape_t{static_cast<ssize_t>(found)});
+            return labels_py;
+        },
+        py::arg("array"),     //
+        py::arg("count") = 10 //
+    );
+
+    hi.def("__len__", &hash_index_w_meta_t::size);
+    hi.def_property_readonly("size", &hash_index_w_meta_t::size);
+    hi.def_property_readonly("connectivity", &hash_index_w_meta_t::connectivity);
+    hi.def_property_readonly("capacity", &hash_index_w_meta_t::capacity);
+
+    hi.def("save", &save_index<hash_index_w_meta_t>, py::arg("path"));
+    hi.def("load", &load_index<hash_index_w_meta_t>, py::arg("path"));
+    hi.def("view", &view_index<hash_index_w_meta_t>, py::arg("path"));
+    hi.def("clear", &clear_index<hash_index_w_meta_t>);
 }

package/python/test.py

@@ -1,7 +1,6 @@
 import numpy as np
 
-import usearch
-
+from usearch.index import Index, SetsIndex
 from numba import cfunc, types, carray
 
 # Showcases how to use Numba to JIT-compile similarity measures for USearch.
@@ -29,13 +28,13 @@ vectors = np.random.uniform(
     0, 0.3, (count_vectors, count_dimensions)).astype(np.float32)
 labels = np.array(range(count_vectors), dtype=np.longlong)
 
-index_udf = usearch.Index(ndim=96, metric_pointer=python_dot.address)
+index_udf = Index(ndim=96, metric_pointer=python_dot.address)
 index_udf.add(labels, vectors, copy=True)
 results = index_udf.search(vectors, 10)
 print('found', results[0].shape, results[1].shape, results[2].shape)
 
 
-index = usearch.Index(ndim=96)
+index = Index(ndim=96)
 index.add(labels, vectors, copy=True)
 assert index.ndim == count_dimensions
 assert index.capacity >= count_vectors
@@ -43,3 +42,10 @@ assert len(index) == count_vectors
 
 results = index.search(vectors, 10)
 print('found', results[0].shape, results[1].shape, results[2].shape)
+
+
+index_sets = SetsIndex()
+index_sets.add(10, np.array([10, 12, 15], dtype=np.uint32))
+index_sets.add(11, np.array([11, 12, 15, 16], dtype=np.uint32))
+results = index_sets.search(np.array([12, 15], dtype=np.uint32), 20)
+print(results)

package/rust/lib.cpp

@@ -24,9 +24,20 @@ Matches Index::search_in_thread(rust::Slice<float const> vector, size_t count, s
     return results;
 }
 
-void Index::add(uint32_t label, rust::Slice<float const> vector) const { return add_in_thread(label, vector, 0); }
+void Index::add(uint32_t label, rust::Slice<float const> vector) const {
+    index_->add(label, vector.data()); //
+}
+
 Matches Index::search(rust::Slice<float const> vector, size_t count) const {
-    return search_in_thread(vector, count, 0);
+    Matches results;
+    results.labels.reserve(count);
+    results.distances.reserve(count);
+    for (size_t i = 0; i != count; ++i)
+        results.labels.push_back(0), results.distances.push_back(0);
+    results.count = index_->search(vector.data(), count, results.labels.data(), results.distances.data());
+    results.labels.truncate(results.count);
+    results.distances.truncate(results.count);
+    return results;
 }
 
 void Index::reserve(size_t capacity) const { index_->reserve(capacity); }

package/setup.py

@@ -9,11 +9,15 @@ compile_args = ['-std=c++11', '-O3']
 if sys.platform == 'darwin':
     compile_args.append('-mmacosx-version-min=10.13')
 
+if sys.platform == 'win32':
+    compile_args = ['/std:c++14', '/O2']
+
+
 ext_modules = [
     Pybind11Extension(
-        'usearch',
+        'usearch.index',
         ['python/lib.cpp'],
-        extra_compile_args=compile_args,
+        extra_compile_args=compile_args
     ),
 ]
 
@@ -30,7 +34,8 @@ setup(
 
     name=__lib_name__,
     version=__version__,
-    packages=find_packages(),
+    packages=['usearch'],
+    package_dir={'usearch': 'python/usearch'},
 
     description='Smaller & Faster Single-File Vector Search Engine from Unum',
     long_description=long_description,

package/simsimd/include/simsimd/simsimd.h

@@ -255,6 +255,21 @@ inline static simsimd_f32_t simsimd_dot_f32x4avx2(simsimd_f32_t const* a, simsim
 #endif
 }
 
+inline static simsimd_f32_t simsimd_dot_f16x16avx512(simsimd_f16_t const* a, simsimd_f16_t const* b, size_t d) {
+#if defined(__AVX512F__)
+    __m512 ab_vec = _mm512_set1_ps(0);
+    for (size_t i = 0; i != d; i += 16)
+        ab_vec = _mm512_fmadd_ps(                     //
+            _mm512_cvtxph_ps(_mm256_loadu_ph(a + i)), //
+            _mm512_cvtxph_ps(_mm256_loadu_ph(b + i)), //
+            ab_vec);
+    return _mm512_reduce_add_ps(ab_vec);
+#else
+    (void)a, (void)b, (void)d;
+    return 0;
+#endif
+}
+
 inline static int32_t simsimd_dot_i8x16avx2(int8_t const* a, int8_t const* b, size_t d) {
 #if defined(__AVX2__)
     __m256i ab_vec = _mm256_set1_epi32(0);