usearch 2.13.4 → 2.14.0

package/README.md

@@ -1,7 +1,7 @@
 <h1 align="center">USearch</h1>
 <h3 align="center">
 Smaller & <a href="https://www.unum.cloud/blog/2023-11-07-scaling-vector-search-with-intel">Faster</a> Single-File<br/>
-Similarity Search Engine for <a href="https://github.com/ashvardanian/simsimd">Vectors</a> & 🔜 <a href="https://github.com/ashvardanian/stringzilla">Texts</a>
+Similarity Search & Clustering Engine for <a href="https://github.com/ashvardanian/simsimd">Vectors</a> & 🔜 <a href="https://github.com/ashvardanian/stringzilla">Texts</a>
 </h3>
 <br/>
 
@@ -71,13 +71,10 @@ Linux • MacOS • Windows • iOS • WebAssembly •
 
 __Technical Insights__ and related articles:
 
-- [Uses Horner's method for polynomial approximations, beating GCC 12 by 119x](https://ashvardanian.com/posts/gcc-12-vs-avx512fp16/).
 - [Uses Arm SVE and x86 AVX-512's masked loads to eliminate tail `for`-loops](https://ashvardanian.com/posts/simsimd-faster-scipy/#tails-of-the-past-the-significance-of-masked-loads).
-- [Uses AVX-512 FP16 for half-precision operations, that few compilers vectorize](https://ashvardanian.com/posts/simsimd-faster-scipy/#the-challenge-of-f16).
-- [Substitutes LibC's `sqrt` calls with bithacks using Jan Kadlec's constant](https://ashvardanian.com/posts/simsimd-faster-scipy/#bonus-section-bypassing-sqrt-and-libc-dependencies).
+- [Uses Horner's method for polynomial approximations, beating GCC 12 by 119x](https://ashvardanian.com/posts/gcc-12-vs-avx512fp16/).
 - [For every language implements a custom separate binding](https://ashvardanian.com/posts/porting-cpp-library-to-ten-languages/).
-- [For Python avoids slow PyBind11, and even `PyArg_ParseTuple` for speed](https://ashvardanian.com/posts/pybind11-cpython-tutorial/).
-- [For JavaScript uses typed arrays and NAPI for zero-copy calls](https://ashvardanian.com/posts/javascript-ai-vector-search/).
+
 
 ## Comparison with FAISS
 
@@ -119,17 +116,15 @@ USearch is compact and broadly compatible without sacrificing performance, prima
 Base functionality is identical to FAISS, and the interface must be familiar if you have ever investigated Approximate Nearest Neighbors search:
 
 ```py
-# pip install numpy usearch
+# pip install usearch
 
 import numpy as np
 from usearch.index import Index
 
-index = Index(ndim=3)
-
-vector = np.array([0.2, 0.6, 0.4])
-index.add(42, vector)
-
-matches = index.search(vector, 10)
+index = Index(ndim=3)               # Default settings for 3D vectors
+vector = np.array([0.2, 0.6, 0.4])  # Can be a matrix for batch operations
+index.add(42, vector)               # Add one or many vectors in parallel
+matches = index.search(vector, 10)  # Find 10 nearest neighbors
 
 assert matches[0].key == 42
 assert matches[0].distance <= 0.001
@@ -137,13 +132,13 @@ assert np.allclose(index[42], vector, atol=0.1) # Ensure high tolerance in mixed
 ```
 
 More settings are always available, and the API is designed to be as flexible as possible.
-The default storage/quantization level is hardware-dependant for efficiency, but `f16` is recommended for most modern CPUs.
+The default storage/quantization level is hardware-dependant for efficiency, but `bf16` is recommended for most modern CPUs.
 
 ```py
 index = Index(
     ndim=3, # Define the number of dimensions in input vectors
     metric='cos', # Choose 'l2sq', 'ip', 'haversine' or other metric, default = 'cos'
-    dtype='f16', # Store as 'f64', 'f32', 'f16', 'i8', 'b1'..., default = None
+    dtype='bf16', # Store as 'f64', 'f32', 'f16', 'i8', 'b1'..., default = None
     connectivity=16, # Optional: Limit number of neighbors per graph node
     expansion_add=128, # Optional: Control the recall of indexing
     expansion_search=64, # Optional: Control the quality of the search
@@ -254,7 +249,7 @@ assert!(
 Training a quantization model and dimension-reduction is a common approach to accelerate vector search.
 Those, however, are only sometimes reliable, can significantly affect the statistical properties of your data, and require regular adjustments if your distribution shifts.
 Instead, we have focused on high-precision arithmetic over low-precision downcasted vectors.
-The same index, and `add` and `search` operations will automatically down-cast or up-cast between `f64_t`, `f32_t`, `f16_t`, `i8_t`, and single-bit representations.
+The same index, and `add` and `search` operations will automatically down-cast or up-cast between `f64_t`, `f32_t`, `f16_t`, `i8_t`, and single-bit `b1x8_t` representations.
 You can use the following command to check, if hardware acceleration is enabled:
 
 ```sh
@@ -264,11 +259,19 @@ $ python -c 'from usearch.index import Index; print(Index(ndim=166, metric="tani
 > ice
 ```
 
+In most cases, it's recommended to use half-precision floating-point numbers on modern hardware.
+When quantization is enabled, the "get"-like functions won't be able to recover the original data, so you may want to replicate the original vectors elsewhere.
+When quantizing to `i8_t` integers, note that it's only valid for cosine-like metrics.
+As part of the quantization process, the vectors are normalized to unit length and later scaled to [-127, 127] range to occupy the full 8-bit range.
+When quantizing to `b1x8_t` single-bit representations, note that it's only valid for binary metrics like Jaccard, Hamming, etc.
+As part of the quantization process, the scalar components greater than zero are set to `true`, and the rest to `false`.
+
+![USearch uint40_t support](https://github.com/unum-cloud/usearch/blob/main/assets/usearch-neighbor-types.png?raw=true)
+
 Using smaller numeric types will save you RAM needed to store the vectors, but you can also compress the neighbors lists forming our proximity graphs.
 By default, 32-bit `uint32_t` is used to enumerate those, which is not enough if you need to address over 4 Billion entries.
 For such cases we provide a custom `uint40_t` type, that will still be 37.5% more space-efficient than the commonly used 8-byte integers, and will scale up to 1 Trillion entries.
 
-![USearch uint40_t support](https://github.com/unum-cloud/usearch/blob/main/assets/usearch-neighbor-types.png?raw=true)
 
 ## `Indexes` for Multi-Index Lookups
 
@@ -541,7 +544,7 @@ doi = {10.5281/zenodo.7949416},
 author = {Vardanian, Ash},
 title = {{USearch by Unum Cloud}},
 url = {https://github.com/unum-cloud/usearch},
-version = {2.13.4},
+version = {2.14.0},
 year = {2023},
 month = oct,
 }

package/c/usearch.h

@@ -72,6 +72,17 @@ USEARCH_EXPORT typedef struct usearch_init_options_t {
     usearch_metric_t metric;
     /**
      *  @brief The scalar kind used for quantization of vector data during indexing.
+     *  In most cases, on modern hardware, it's recommended to use half-precision floating-point numbers.
+     *  When quantization is enabled, the "get"-like functions won't be able to recover the original data,
+     *  so you may want to replicate the original vectors elsewhere.
+     *
+     *  Quantizing to integers is also possible, but it's important to note that it's only valid for cosine-like
+     *  metrics. As part of the quantization process, the vectors are normalized to unit length and later scaled
+     *  to @b [-127,127] range to occupy the full 8-bit range.
+     *
+     *  Quantizing to 1-bit booleans is also possible, but it's only valid for binary metrics like Jaccard, Hamming,
+     *  etc. As part of the quantization process, the scalar components greater than zero are set to `true`, and the
+     *  rest to `false`.
      */
     usearch_scalar_kind_t quantization;
     /**

package/include/usearch/index.hpp

@@ -8,8 +8,8 @@
 #define UNUM_USEARCH_HPP
 
 #define USEARCH_VERSION_MAJOR 2
-#define USEARCH_VERSION_MINOR 13
-#define USEARCH_VERSION_PATCH 4
+#define USEARCH_VERSION_MINOR 14
+#define USEARCH_VERSION_PATCH 0
 
 // Inferring C++ version
 // https://stackoverflow.com/a/61552074
@@ -2870,16 +2870,16 @@ class index_gt {
         result.computed_distances = context.computed_distances;
         result.visited_members = context.iteration_cycles;
 
-        // If we are updating the entry node itself, it won't contain any neighbors,
-        // so we should traverse a level down to find the closest match.
-        if (updated_node_level == max_level_copy)
-            updated_node_level--;
-
         // Go down the level, tracking only the closest match;
         // It may even be equal to the `updated_slot`
-        compressed_slot_t closest_slot = search_for_one_( //
-            value, metric, prefetch,                      //
-            entry_slot_copy, max_level_copy, updated_node_level, context);
+        compressed_slot_t closest_slot =
+            // If we are updating the entry node itself, it won't contain any neighbors,
+            // so we should traverse a level down to find the closest match.
+            updated_node_level == max_level_copy //
+                ? entry_slot_copy
+                : search_for_one_(             //
+                      value, metric, prefetch, //
+                      entry_slot_copy, max_level_copy, updated_node_level, context);
 
         // From `updated_node_level` down - perform proper extensive search
         for (level_t level = (std::min)(updated_node_level, max_level_copy); level >= 0; --level) {
@@ -2943,7 +2943,7 @@ class index_gt {
             config.expansion = default_expansion_search();
 
         // Using references is cleaner, but would result in UBSan false positives
-        context_t* context_ptr = contexts_.data() + config.thread;
+        context_t* context_ptr = contexts_.data() ? contexts_.data() + config.thread : nullptr;
         top_candidates_t* top_ptr = context_ptr ? &context_ptr->top_candidates : nullptr;
         search_result_t result{*this, top_ptr};
         if (!nodes_count_.load(std::memory_order_relaxed))
@@ -3697,6 +3697,7 @@ class index_gt {
     inline neighbors_ref_t neighbors_base_(node_t node) const noexcept { return {node.neighbors_tape()}; }
 
     inline neighbors_ref_t neighbors_non_base_(node_t node, level_t level) const noexcept {
+        usearch_assert_m(level > 0 && level <= node.level(), "Linking to missing level");
         return {node.neighbors_tape() + pre_.neighbors_base_bytes + (level - 1) * pre_.neighbors_bytes};
     }
 

package/include/usearch/index_dense.hpp

@@ -95,7 +95,7 @@ struct index_dense_head_result_t {
  *  @brief  Configuration settings for the construction of dense
  *          equidimensional vector indexes.
  *
- *  Unlike the underlying `index_gt` class, incroporates the
+ *  Unlike the underlying `index_gt` class, incorporates the
  *  `::expansion_add` and `::expansion_search` parameters passed
  *  separately for the lower-level engine.
  */
@@ -207,8 +207,11 @@ struct index_dense_metadata_result_t {
     }
 };
 
+/**
+ *  @brief  Fixes serialized scalar-kind codes for pre-v2.10 versions, until we can upgrade to v3.
+ *          The old enum `scalar_kind_t` is defined without explicit constants from 0.
+ */
 inline scalar_kind_t convert_pre_2_10_scalar_kind(scalar_kind_t scalar_kind) noexcept {
-    // The old enum `scalar_kind_t` is defined without explicit constants from 0
     switch (static_cast<std::underlying_type<scalar_kind_t>::type>(scalar_kind)) {
     case 0: return scalar_kind_t::unknown_k;
     case 1: return scalar_kind_t::b1x8_k;
@@ -229,13 +232,15 @@ inline scalar_kind_t convert_pre_2_10_scalar_kind(scalar_kind_t scalar_kind) noe
     }
 }
 
+/**
+ *  @brief  Fixes the metadata for pre-v2.10 versions, until we can upgrade to v3.
+ *          Originates from: https://github.com/unum-cloud/usearch/issues/423
+ */
 inline void fix_pre_2_10_metadata(index_dense_head_t& head) {
-    // deal with incompatibility with pre-2.10 versions: https://github.com/unum-cloud/usearch/issues/423
     if (head.version_major == 2 && head.version_minor < 10) {
         head.kind_scalar = convert_pre_2_10_scalar_kind(head.kind_scalar);
         head.kind_key = convert_pre_2_10_scalar_kind(head.kind_key);
         head.kind_compressed_slot = convert_pre_2_10_scalar_kind(head.kind_compressed_slot);
-        // update minor versions so that no repeated conversion occurs
         head.version_minor = 10;
         head.version_patch = 0;
     }
@@ -316,7 +321,7 @@ inline index_dense_metadata_result_t index_dense_metadata_from_path(char const*
 /**
  *  @brief  Extracts metadata from a pre-constructed index serialized into an in-memory buffer.
  */
-inline index_dense_metadata_result_t index_dense_metadata_from_buffer(memory_mapped_file_t file,
+inline index_dense_metadata_result_t index_dense_metadata_from_buffer(memory_mapped_file_t const& file,
                                                                       std::size_t offset = 0) noexcept {
     index_dense_metadata_result_t result;
 
@@ -324,7 +329,7 @@ inline index_dense_metadata_result_t index_dense_metadata_from_buffer(memory_map
     if (offset + sizeof(index_dense_head_buffer_t) >= file.size())
         return result.failed("End of file reached!");
 
-    byte_t* const file_data = file.data() + offset;
+    byte_t const* file_data = file.data() + offset;
     std::size_t const file_size = file.size() - offset;
     std::memcpy(&result.head_buffer, file_data, sizeof(index_dense_head_buffer_t));
 
@@ -2235,6 +2240,7 @@ class index_dense_gt {
         case scalar_kind_t::f64_k: return make_casts_<f64_t>();
         case scalar_kind_t::f32_k: return make_casts_<f32_t>();
         case scalar_kind_t::f16_k: return make_casts_<f16_t>();
+        case scalar_kind_t::bf16_k: return make_casts_<bf16_t>();
         case scalar_kind_t::i8_k: return make_casts_<i8_t>();
         case scalar_kind_t::b1x8_k: return make_casts_<b1x8_t>();
         default: return {};

package/include/usearch/index_plugins.hpp

@@ -70,20 +70,26 @@ struct uuid_t {
 };
 
 class f16_bits_t;
-class i8_converted_t;
+class bf16_bits_t;
 
 #if !USEARCH_USE_FP16LIB
 #if USEARCH_USE_SIMSIMD
 using f16_native_t = simsimd_f16_t;
+using bf16_native_t = simsimd_bf16_t;
 #elif defined(USEARCH_DEFINED_ARM)
 using f16_native_t = __fp16;
+using bf16_native_t = __fp16; // No better choice on most compilers!
 #else
 using f16_native_t = _Float16;
+using bf16_native_t = _Float16; // No better choice on most compilers!
 #endif
 using f16_t = f16_native_t;
+using bf16_t = bf16_native_t;
 #else
 using f16_native_t = void;
+using bf16_native_t = void;
 using f16_t = f16_bits_t;
+using bf16_t = bf16_bits_t;
 #endif
 
 using f64_t = double;
@@ -99,6 +105,9 @@ using i32_t = std::int32_t;
 using i16_t = std::int16_t;
 using i8_t = std::int8_t;
 
+/**
+ *  @brief  Enumerates the most commonly used distance metrics, mostly for dense vector representations.
+ */
 enum class metric_kind_t : std::uint8_t {
     unknown_k = 0,
     // Classics:
@@ -111,19 +120,26 @@ enum class metric_kind_t : std::uint8_t {
     haversine_k = 'h',
     divergence_k = 'd',
 
-    // Sets:
-    jaccard_k = 'j',
+    // Dense Sets:
     hamming_k = 'b',
     tanimoto_k = 't',
     sorensen_k = 's',
+
+    // Sparse Sets:
+    jaccard_k = 'j',
 };
 
+/**
+ *  @brief  Enumerates the most commonly used scalar types, mostly for dense vector representations.
+ *          Doesn't include logical types, like complex numbers or quaternions.
+ */
 enum class scalar_kind_t : std::uint8_t {
     unknown_k = 0,
     // Custom:
     b1x8_k = 1,
     u40_k = 2,
     uuid_k = 3,
+    bf16_k = 4,
     // Common:
     f64_k = 10,
     f32_k = 11,
@@ -159,6 +175,8 @@ template <typename scalar_at> scalar_kind_t scalar_kind() noexcept {
         return scalar_kind_t::f32_k;
     if (std::is_same<scalar_at, f16_t>())
         return scalar_kind_t::f16_k;
+    if (std::is_same<scalar_at, bf16_t>())
+        return scalar_kind_t::bf16_k;
     if (std::is_same<scalar_at, i8_t>())
         return scalar_kind_t::i8_k;
     if (std::is_same<scalar_at, u64_t>())
@@ -198,33 +216,66 @@ inline bool str_equals(char const* begin, std::size_t len, char const* other_beg
 
 inline std::size_t bits_per_scalar(scalar_kind_t scalar_kind) noexcept {
     switch (scalar_kind) {
+    case scalar_kind_t::uuid_k: return 128;
+    case scalar_kind_t::u40_k: return 40;
+    case scalar_kind_t::bf16_k: return 16;
+    case scalar_kind_t::b1x8_k: return 1;
+    case scalar_kind_t::u64_k: return 64;
+    case scalar_kind_t::i64_k: return 64;
     case scalar_kind_t::f64_k: return 64;
+    case scalar_kind_t::u32_k: return 32;
+    case scalar_kind_t::i32_k: return 32;
     case scalar_kind_t::f32_k: return 32;
+    case scalar_kind_t::u16_k: return 16;
+    case scalar_kind_t::i16_k: return 16;
     case scalar_kind_t::f16_k: return 16;
+    case scalar_kind_t::u8_k: return 8;
     case scalar_kind_t::i8_k: return 8;
-    case scalar_kind_t::b1x8_k: return 1;
+    case scalar_kind_t::f8_k: return 8;
     default: return 0;
     }
 }
 
 inline std::size_t bits_per_scalar_word(scalar_kind_t scalar_kind) noexcept {
     switch (scalar_kind) {
+    case scalar_kind_t::uuid_k: return 128;
+    case scalar_kind_t::u40_k: return 40;
+    case scalar_kind_t::bf16_k: return 16;
+    case scalar_kind_t::b1x8_k: return 8;
+    case scalar_kind_t::u64_k: return 64;
+    case scalar_kind_t::i64_k: return 64;
     case scalar_kind_t::f64_k: return 64;
+    case scalar_kind_t::u32_k: return 32;
+    case scalar_kind_t::i32_k: return 32;
     case scalar_kind_t::f32_k: return 32;
+    case scalar_kind_t::u16_k: return 16;
+    case scalar_kind_t::i16_k: return 16;
     case scalar_kind_t::f16_k: return 16;
+    case scalar_kind_t::u8_k: return 8;
     case scalar_kind_t::i8_k: return 8;
-    case scalar_kind_t::b1x8_k: return 8;
+    case scalar_kind_t::f8_k: return 8;
     default: return 0;
     }
 }
 
 inline char const* scalar_kind_name(scalar_kind_t scalar_kind) noexcept {
     switch (scalar_kind) {
+    case scalar_kind_t::uuid_k: return "uuid";
+    case scalar_kind_t::u40_k: return "u40";
+    case scalar_kind_t::bf16_k: return "bf16";
+    case scalar_kind_t::b1x8_k: return "b1x8";
+    case scalar_kind_t::u64_k: return "u64";
+    case scalar_kind_t::i64_k: return "i64";
+    case scalar_kind_t::f64_k: return "f64";
+    case scalar_kind_t::u32_k: return "u32";
+    case scalar_kind_t::i32_k: return "i32";
     case scalar_kind_t::f32_k: return "f32";
+    case scalar_kind_t::u16_k: return "u16";
+    case scalar_kind_t::i16_k: return "i16";
     case scalar_kind_t::f16_k: return "f16";
-    case scalar_kind_t::f64_k: return "f64";
+    case scalar_kind_t::u8_k: return "u8";
     case scalar_kind_t::i8_k: return "i8";
-    case scalar_kind_t::b1x8_k: return "b1x8";
+    case scalar_kind_t::f8_k: return "f8";
     default: return "";
     }
 }
@@ -242,8 +293,8 @@ inline char const* metric_kind_name(metric_kind_t metric) noexcept {
     case metric_kind_t::hamming_k: return "hamming";
     case metric_kind_t::tanimoto_k: return "tanimoto";
     case metric_kind_t::sorensen_k: return "sorensen";
+    default: return "";
     }
-    return "";
 }
 inline expected_gt<scalar_kind_t> scalar_kind_from_name(char const* name, std::size_t len) {
     expected_gt<scalar_kind_t> parsed;
@@ -253,10 +304,14 @@ inline expected_gt<scalar_kind_t> scalar_kind_from_name(char const* name, std::s
         parsed.result = scalar_kind_t::f64_k;
     else if (str_equals(name, len, "f16"))
         parsed.result = scalar_kind_t::f16_k;
+    else if (str_equals(name, len, "bf16"))
+        parsed.result = scalar_kind_t::bf16_k;
     else if (str_equals(name, len, "i8"))
         parsed.result = scalar_kind_t::i8_k;
+    else if (str_equals(name, len, "b1"))
+        parsed.result = scalar_kind_t::b1x8_k;
     else
-        parsed.failed("Unknown type, choose: f32, f16, f64, i8");
+        parsed.failed("Unknown type, choose: f64, f32, f16, bf16, i8, b1");
     return parsed;
 }
 
@@ -289,11 +344,13 @@ inline expected_gt<metric_kind_t> metric_from_name(char const* name, std::size_t
                       "tanimoto, sorensen");
     return parsed;
 }
-
 inline expected_gt<metric_kind_t> metric_from_name(char const* name) {
     return metric_from_name(name, std::strlen(name));
 }
 
+/**
+ *  @brief Convenience function to upcast a half-precision floating point number to a single-precision one.
+ */
 inline float f16_to_f32(std::uint16_t u16) noexcept {
 #if !USEARCH_USE_FP16LIB
     f16_native_t f16;
@@ -304,6 +361,9 @@ inline float f16_to_f32(std::uint16_t u16) noexcept {
 #endif
 }
 
+/**
+ *  @brief Convenience function to downcast a single-precision floating point number to a half-precision one.
+ */
 inline std::uint16_t f32_to_f16(float f32) noexcept {
 #if !USEARCH_USE_FP16LIB
     f16_native_t f16 = f16_native_t(f32);
@@ -315,6 +375,36 @@ inline std::uint16_t f32_to_f16(float f32) noexcept {
 #endif
 }
 
+/**
+ *  @brief Convenience function to upcast a brain-floating point number to a single-precision one.
+ *  https://github.com/ashvardanian/SimSIMD/blob/ff51434d90c66f916e94ff05b24530b127aa4cff/include/simsimd/types.h#L395-L410
+ */
+inline float bf16_to_f32(std::uint16_t u16) noexcept {
+    union float_or_unsigned_int_t {
+        float f;
+        unsigned int i;
+    };
+    union float_or_unsigned_int_t result_union;
+    result_union.i = u16 << 16; // Zero extends the mantissa
+    return result_union.f;
+}
+
+/**
+ *  @brief Convenience function to downcast a single-precision floating point number to a brain-floating point one.
+ *  https://github.com/ashvardanian/SimSIMD/blob/ff51434d90c66f916e94ff05b24530b127aa4cff/include/simsimd/types.h#L412-L425
+ */
+inline std::uint16_t f32_to_bf16(float f32) noexcept {
+    union float_or_unsigned_int_t {
+        float f;
+        unsigned int i;
+    };
+    union float_or_unsigned_int_t value;
+    value.f = f32;
+    value.i >>= 16;
+    value.i &= 0xFFFF;
+    return (unsigned short)value.i;
+}
+
 /**
  *  @brief  Numeric type for the IEEE 754 half-precision floating point.
  *          If hardware support isn't available, falls back to a hardware
@@ -333,8 +423,7 @@ class f16_bits_t {
     inline operator float() const noexcept { return f16_to_f32(uint16_); }
     inline explicit operator bool() const noexcept { return f16_to_f32(uint16_) > 0.5f; }
 
-    inline f16_bits_t(i8_converted_t) noexcept;
-    inline f16_bits_t(std::int8_t v) noexcept : uint16_(v) {}
+    inline f16_bits_t(std::int8_t v) noexcept : uint16_(f32_to_f16(v)) {}
     inline f16_bits_t(bool v) noexcept : uint16_(f32_to_f16(v)) {}
     inline f16_bits_t(float v) noexcept : uint16_(f32_to_f16(v)) {}
     inline f16_bits_t(double v) noexcept : uint16_(f32_to_f16(static_cast<float>(v))) {}
@@ -345,14 +434,14 @@ class f16_bits_t {
     inline f16_bits_t operator-(f16_bits_t other) const noexcept { return {float(*this) - float(other)}; }
     inline f16_bits_t operator*(f16_bits_t other) const noexcept { return {float(*this) * float(other)}; }
     inline f16_bits_t operator/(f16_bits_t other) const noexcept { return {float(*this) / float(other)}; }
-    inline f16_bits_t operator+(float other) const noexcept { return {float(*this) + other}; }
-    inline f16_bits_t operator-(float other) const noexcept { return {float(*this) - other}; }
-    inline f16_bits_t operator*(float other) const noexcept { return {float(*this) * other}; }
-    inline f16_bits_t operator/(float other) const noexcept { return {float(*this) / other}; }
-    inline f16_bits_t operator+(double other) const noexcept { return {float(*this) + other}; }
-    inline f16_bits_t operator-(double other) const noexcept { return {float(*this) - other}; }
-    inline f16_bits_t operator*(double other) const noexcept { return {float(*this) * other}; }
-    inline f16_bits_t operator/(double other) const noexcept { return {float(*this) / other}; }
+    inline float operator+(float other) const noexcept { return float(*this) + other; }
+    inline float operator-(float other) const noexcept { return float(*this) - other; }
+    inline float operator*(float other) const noexcept { return float(*this) * other; }
+    inline float operator/(float other) const noexcept { return float(*this) / other; }
+    inline double operator+(double other) const noexcept { return float(*this) + other; }
+    inline double operator-(double other) const noexcept { return float(*this) - other; }
+    inline double operator*(double other) const noexcept { return float(*this) * other; }
+    inline double operator/(double other) const noexcept { return float(*this) / other; }
 
     inline f16_bits_t& operator+=(float v) noexcept {
         uint16_ = f32_to_f16(v + f16_to_f32(uint16_));
@@ -375,6 +464,65 @@ class f16_bits_t {
     }
 };
 
+/**
+ *  @brief  Numeric type for brain-floating point half-precision floating point.
+ *          If hardware support isn't available, falls back to a hardware
+ *          agnostic in-software implementation.
+ */
+class bf16_bits_t {
+    std::uint16_t uint16_{};
+
+  public:
+    inline bf16_bits_t() noexcept : uint16_(0) {}
+    inline bf16_bits_t(bf16_bits_t&&) = default;
+    inline bf16_bits_t& operator=(bf16_bits_t&&) = default;
+    inline bf16_bits_t(bf16_bits_t const&) = default;
+    inline bf16_bits_t& operator=(bf16_bits_t const&) = default;
+
+    inline operator float() const noexcept { return bf16_to_f32(uint16_); }
+    inline explicit operator bool() const noexcept { return bf16_to_f32(uint16_) > 0.5f; }
+
+    inline bf16_bits_t(std::int8_t v) noexcept : uint16_(f32_to_bf16(v)) {}
+    inline bf16_bits_t(bool v) noexcept : uint16_(f32_to_bf16(v)) {}
+    inline bf16_bits_t(float v) noexcept : uint16_(f32_to_bf16(v)) {}
+    inline bf16_bits_t(double v) noexcept : uint16_(f32_to_bf16(static_cast<float>(v))) {}
+
+    inline bool operator<(bf16_bits_t const& other) const noexcept { return float(*this) < float(other); }
+
+    inline bf16_bits_t operator+(bf16_bits_t other) const noexcept { return {float(*this) + float(other)}; }
+    inline bf16_bits_t operator-(bf16_bits_t other) const noexcept { return {float(*this) - float(other)}; }
+    inline bf16_bits_t operator*(bf16_bits_t other) const noexcept { return {float(*this) * float(other)}; }
+    inline bf16_bits_t operator/(bf16_bits_t other) const noexcept { return {float(*this) / float(other)}; }
+    inline float operator+(float other) const noexcept { return float(*this) + other; }
+    inline float operator-(float other) const noexcept { return float(*this) - other; }
+    inline float operator*(float other) const noexcept { return float(*this) * other; }
+    inline float operator/(float other) const noexcept { return float(*this) / other; }
+    inline double operator+(double other) const noexcept { return float(*this) + other; }
+    inline double operator-(double other) const noexcept { return float(*this) - other; }
+    inline double operator*(double other) const noexcept { return float(*this) * other; }
+    inline double operator/(double other) const noexcept { return float(*this) / other; }
+
+    inline bf16_bits_t& operator+=(float v) noexcept {
+        uint16_ = f32_to_bf16(v + bf16_to_f32(uint16_));
+        return *this;
+    }
+
+    inline bf16_bits_t& operator-=(float v) noexcept {
+        uint16_ = f32_to_bf16(v - bf16_to_f32(uint16_));
+        return *this;
+    }
+
+    inline bf16_bits_t& operator*=(float v) noexcept {
+        uint16_ = f32_to_bf16(v * bf16_to_f32(uint16_));
+        return *this;
+    }
+
+    inline bf16_bits_t& operator/=(float v) noexcept {
+        uint16_ = f32_to_bf16(v / bf16_to_f32(uint16_));
+        return *this;
+    }
+};
+
 /**
  *  @brief  An STL-based executor or a "thread-pool" for parallel execution.
  *          Isn't efficient for small batches, as it recreates the threads on every call.
@@ -907,7 +1055,7 @@ template <> struct cast_gt<b1x8_t, b1x8_t> {
     static bool try_(byte_t const*, std::size_t, byte_t*) noexcept { return false; }
 };
 
-template <typename from_scalar_at> struct cast_gt<from_scalar_at, b1x8_t> {
+template <typename from_scalar_at> struct cast_to_b1x8_gt {
     inline static bool try_(byte_t const* input, std::size_t dim, byte_t* output) noexcept {
         from_scalar_at const* typed_input = reinterpret_cast<from_scalar_at const*>(input);
         unsigned char* typed_output = reinterpret_cast<unsigned char*>(output);
@@ -928,7 +1076,7 @@ template <typename from_scalar_at> struct cast_gt<from_scalar_at, b1x8_t> {
     }
 };
 
-template <typename to_scalar_at> struct cast_gt<b1x8_t, to_scalar_at> {
+template <typename to_scalar_at> struct cast_from_b1x8_gt {
     static bool try_(byte_t const* input, std::size_t dim, byte_t* output) noexcept {
         unsigned char const* typed_input = reinterpret_cast<unsigned char const*>(input);
         to_scalar_at* typed_output = reinterpret_cast<to_scalar_at*>(output);
@@ -940,53 +1088,59 @@ template <typename to_scalar_at> struct cast_gt<b1x8_t, to_scalar_at> {
     }
 };
 
-/**
- *  @brief  Numeric type for uniformly-distributed floating point
- *          values within [-1,1] range, quantized to integers [-100,100].
- */
-class i8_converted_t {
-    std::int8_t int8_{};
+template <typename from_scalar_at> struct cast_to_i8_gt {
+    inline static bool try_(byte_t const* input, std::size_t dim, byte_t* output) noexcept {
+        from_scalar_at const* typed_input = reinterpret_cast<from_scalar_at const*>(input);
+        std::int8_t* typed_output = reinterpret_cast<std::int8_t*>(output);
+        // Unlike other casting mechanisms, switching to small range integers is a two step procedure.
+        // First we want to estimate the magnitude of the vector to scale into [-1.0, 1.0] interval,
+        // instead of clamping. And then we scale the values into the [-127, 127] range.
+        // ! This makes an assumption, that the distance metric is dot-product-like, which may not
+        // ! be true in many cases, so it's recommended to avoid automatic casting from floats to
+        // ! integers.
+        double magnitude = 0.0;
+        for (std::size_t i = 0; i != dim; ++i)
+            magnitude += (double)typed_input[i] * (double)typed_input[i];
+        magnitude = std::sqrt(magnitude);
+        for (std::size_t i = 0; i != dim; ++i)
+            typed_output[i] =
+                static_cast<std::int8_t>(usearch::clamp<double>(typed_input[i] * 127.0 / magnitude, -127.0, 127.0));
+        return true;
+    }
+};
 
-  public:
-    constexpr static f32_t divisor_k = 100.f;
-    constexpr static std::int8_t min_k = -100;
-    constexpr static std::int8_t max_k = 100;
-
-    inline i8_converted_t() noexcept : int8_(0) {}
-    inline i8_converted_t(bool v) noexcept : int8_(v ? max_k : 0) {}
-
-    inline i8_converted_t(i8_converted_t&&) = default;
-    inline i8_converted_t& operator=(i8_converted_t&&) = default;
-    inline i8_converted_t(i8_converted_t const&) = default;
-    inline i8_converted_t& operator=(i8_converted_t const&) = default;
-
-    inline operator f16_t() const noexcept { return static_cast<f16_t>(f32_t(int8_) / divisor_k); }
-    inline operator f32_t() const noexcept { return f32_t(int8_) / divisor_k; }
-    inline operator f64_t() const noexcept { return f64_t(int8_) / divisor_k; }
-    inline explicit operator bool() const noexcept { return int8_ > (max_k / 2); }
-    inline explicit operator std::int8_t() const noexcept { return int8_; }
-    inline explicit operator std::int16_t() const noexcept { return int8_; }
-    inline explicit operator std::int32_t() const noexcept { return int8_; }
-    inline explicit operator std::int64_t() const noexcept { return int8_; }
-
-    // Even with `f16_t` arguments we may want to use `f32_t` for clamping and comparions.
-    inline i8_converted_t(f16_t v)
-        : int8_(static_cast<std::int8_t>(usearch::clamp<f32_t>(v * divisor_k, min_k, max_k))) {}
-    inline i8_converted_t(f32_t v)
-        : int8_(static_cast<std::int8_t>(usearch::clamp<f32_t>(v * divisor_k, min_k, max_k))) {}
-    inline i8_converted_t(f64_t v)
-        : int8_(static_cast<std::int8_t>(usearch::clamp<f64_t>(v * divisor_k, min_k, max_k))) {}
+template <typename to_scalar_at> struct cast_from_i8_gt {
+    static bool try_(byte_t const* input, std::size_t dim, byte_t* output) noexcept {
+        std::int8_t const* typed_input = reinterpret_cast<std::int8_t const*>(input);
+        to_scalar_at* typed_output = reinterpret_cast<to_scalar_at*>(output);
+        for (std::size_t i = 0; i != dim; ++i)
+            typed_output[i] = static_cast<to_scalar_at>(typed_input[i]) / 127.f;
+        return true;
+    }
 };
 
-f16_bits_t::f16_bits_t(i8_converted_t v) noexcept : uint16_(f32_to_f16(v)) {}
+template <> struct cast_gt<i8_t, f16_t> : public cast_from_i8_gt<f16_t> {};
+template <> struct cast_gt<i8_t, bf16_t> : public cast_from_i8_gt<bf16_t> {};
+template <> struct cast_gt<i8_t, f32_t> : public cast_from_i8_gt<f32_t> {};
+template <> struct cast_gt<i8_t, f64_t> : public cast_from_i8_gt<f64_t> {};
+
+template <> struct cast_gt<f16_t, i8_t> : public cast_to_i8_gt<f16_t> {};
+template <> struct cast_gt<bf16_t, i8_t> : public cast_to_i8_gt<bf16_t> {};
+template <> struct cast_gt<f32_t, i8_t> : public cast_to_i8_gt<f32_t> {};
+template <> struct cast_gt<f64_t, i8_t> : public cast_to_i8_gt<f64_t> {};
 
-template <> struct cast_gt<i8_t, f16_t> : public cast_gt<i8_converted_t, f16_t> {};
-template <> struct cast_gt<i8_t, f32_t> : public cast_gt<i8_converted_t, f32_t> {};
-template <> struct cast_gt<i8_t, f64_t> : public cast_gt<i8_converted_t, f64_t> {};
+template <> struct cast_gt<b1x8_t, f16_t> : public cast_from_b1x8_gt<f16_t> {};
+template <> struct cast_gt<b1x8_t, bf16_t> : public cast_from_b1x8_gt<bf16_t> {};
+template <> struct cast_gt<b1x8_t, f32_t> : public cast_from_b1x8_gt<f32_t> {};
+template <> struct cast_gt<b1x8_t, f64_t> : public cast_from_b1x8_gt<f64_t> {};
 
-template <> struct cast_gt<f16_t, i8_t> : public cast_gt<f16_t, i8_converted_t> {};
-template <> struct cast_gt<f32_t, i8_t> : public cast_gt<f32_t, i8_converted_t> {};
-template <> struct cast_gt<f64_t, i8_t> : public cast_gt<f64_t, i8_converted_t> {};
+template <> struct cast_gt<f16_t, b1x8_t> : public cast_to_b1x8_gt<f16_t> {};
+template <> struct cast_gt<bf16_t, b1x8_t> : public cast_to_b1x8_gt<bf16_t> {};
+template <> struct cast_gt<f32_t, b1x8_t> : public cast_to_b1x8_gt<f32_t> {};
+template <> struct cast_gt<f64_t, b1x8_t> : public cast_to_b1x8_gt<f64_t> {};
+
+template <> struct cast_gt<b1x8_t, i8_t> : public cast_from_b1x8_gt<i8_t> {};
+template <> struct cast_gt<i8_t, b1x8_t> : public cast_to_b1x8_gt<i8_t> {};
 
 /*  Don't complain if the vectorization of the inner loops fails:
  *
@@ -1563,6 +1717,7 @@ class metric_punned_t {
         case scalar_kind_t::f32_k: datatype = simsimd_datatype_f32_k; break;
         case scalar_kind_t::f64_k: datatype = simsimd_datatype_f64_k; break;
         case scalar_kind_t::f16_k: datatype = simsimd_datatype_f16_k; break;
+        case scalar_kind_t::bf16_k: datatype = simsimd_datatype_bf16_k; break;
         case scalar_kind_t::i8_k: datatype = simsimd_datatype_i8_k; break;
         case scalar_kind_t::b1x8_k: datatype = simsimd_datatype_b8_k; break;
         default: break;
@@ -1615,9 +1770,10 @@ class metric_punned_t {
         switch (metric_kind_) {
         case metric_kind_t::ip_k: {
             switch (scalar_kind_) {
-            case scalar_kind_t::f32_k: metric_ptr_ = (uptr_t)&equidimensional_<metric_ip_gt<f32_t>>; break;
-            case scalar_kind_t::f16_k: metric_ptr_ = (uptr_t)&equidimensional_<metric_ip_gt<f16_t, f32_t>>; break;
+            case scalar_kind_t::bf16_k: metric_ptr_ = (uptr_t)&equidimensional_<metric_ip_gt<bf16_t, f32_t>>; break;
             case scalar_kind_t::i8_k: metric_ptr_ = (uptr_t)&equidimensional_<metric_ip_gt<i8_t, f32_t>>; break;
+            case scalar_kind_t::f16_k: metric_ptr_ = (uptr_t)&equidimensional_<metric_ip_gt<f16_t, f32_t>>; break;
+            case scalar_kind_t::f32_k: metric_ptr_ = (uptr_t)&equidimensional_<metric_ip_gt<f32_t>>; break;
             case scalar_kind_t::f64_k: metric_ptr_ = (uptr_t)&equidimensional_<metric_ip_gt<f64_t>>; break;
             default: metric_ptr_ = 0; break;
             }
@@ -1625,9 +1781,10 @@ class metric_punned_t {
         }
         case metric_kind_t::cos_k: {
             switch (scalar_kind_) {
-            case scalar_kind_t::f32_k: metric_ptr_ = (uptr_t)&equidimensional_<metric_cos_gt<f32_t>>; break;
-            case scalar_kind_t::f16_k: metric_ptr_ = (uptr_t)&equidimensional_<metric_cos_gt<f16_t, f32_t>>; break;
+            case scalar_kind_t::bf16_k: metric_ptr_ = (uptr_t)&equidimensional_<metric_cos_gt<bf16_t, f32_t>>; break;
             case scalar_kind_t::i8_k: metric_ptr_ = (uptr_t)&equidimensional_<metric_cos_gt<i8_t, f32_t>>; break;
+            case scalar_kind_t::f16_k: metric_ptr_ = (uptr_t)&equidimensional_<metric_cos_gt<f16_t, f32_t>>; break;
+            case scalar_kind_t::f32_k: metric_ptr_ = (uptr_t)&equidimensional_<metric_cos_gt<f32_t>>; break;
             case scalar_kind_t::f64_k: metric_ptr_ = (uptr_t)&equidimensional_<metric_cos_gt<f64_t>>; break;
             default: metric_ptr_ = 0; break;
             }
@@ -1635,9 +1792,10 @@ class metric_punned_t {
         }
         case metric_kind_t::l2sq_k: {
             switch (scalar_kind_) {
-            case scalar_kind_t::f32_k: metric_ptr_ = (uptr_t)&equidimensional_<metric_l2sq_gt<f32_t>>; break;
-            case scalar_kind_t::f16_k: metric_ptr_ = (uptr_t)&equidimensional_<metric_l2sq_gt<f16_t, f32_t>>; break;
+            case scalar_kind_t::bf16_k: metric_ptr_ = (uptr_t)&equidimensional_<metric_l2sq_gt<bf16_t, f32_t>>; break;
             case scalar_kind_t::i8_k: metric_ptr_ = (uptr_t)&equidimensional_<metric_l2sq_gt<i8_t, f32_t>>; break;
+            case scalar_kind_t::f16_k: metric_ptr_ = (uptr_t)&equidimensional_<metric_l2sq_gt<f16_t, f32_t>>; break;
+            case scalar_kind_t::f32_k: metric_ptr_ = (uptr_t)&equidimensional_<metric_l2sq_gt<f32_t>>; break;
             case scalar_kind_t::f64_k: metric_ptr_ = (uptr_t)&equidimensional_<metric_l2sq_gt<f64_t>>; break;
             default: metric_ptr_ = 0; break;
             }
@@ -1645,6 +1803,9 @@ class metric_punned_t {
         }
         case metric_kind_t::pearson_k: {
             switch (scalar_kind_) {
+            case scalar_kind_t::bf16_k:
+                metric_ptr_ = (uptr_t)&equidimensional_<metric_pearson_gt<bf16_t, f32_t>>;
+                break;
             case scalar_kind_t::i8_k: metric_ptr_ = (uptr_t)&equidimensional_<metric_pearson_gt<i8_t, f32_t>>; break;
             case scalar_kind_t::f16_k: metric_ptr_ = (uptr_t)&equidimensional_<metric_pearson_gt<f16_t, f32_t>>; break;
             case scalar_kind_t::f32_k: metric_ptr_ = (uptr_t)&equidimensional_<metric_pearson_gt<f32_t>>; break;
@@ -1655,7 +1816,8 @@ class metric_punned_t {
         }
         case metric_kind_t::haversine_k: {
             switch (scalar_kind_) {
-            case scalar_kind_t::f16_k: metric_ptr_ = 0; break; //< Having half-precision 2D coordinates is a bit silly.
+            case scalar_kind_t::bf16_k: metric_ptr_ = 0; break; //< Half-precision 2D vectors are silly.
+            case scalar_kind_t::f16_k: metric_ptr_ = 0; break;  //< Half-precision 2D vectors are silly.
             case scalar_kind_t::f32_k: metric_ptr_ = (uptr_t)&equidimensional_<metric_haversine_gt<f32_t>>; break;
             case scalar_kind_t::f64_k: metric_ptr_ = (uptr_t)&equidimensional_<metric_haversine_gt<f64_t>>; break;
             default: metric_ptr_ = 0; break;
@@ -1664,6 +1826,9 @@ class metric_punned_t {
         }
         case metric_kind_t::divergence_k: {
             switch (scalar_kind_) {
+            case scalar_kind_t::bf16_k:
+                metric_ptr_ = (uptr_t)&equidimensional_<metric_divergence_gt<bf16_t, f32_t>>;
+                break;
             case scalar_kind_t::f16_k:
                 metric_ptr_ = (uptr_t)&equidimensional_<metric_divergence_gt<f16_t, f32_t>>;
                 break;

package/javascript/dist/cjs/usearch.d.ts

@@ -29,6 +29,7 @@ export declare enum ScalarKind {
     F32 = "f32",
     F64 = "f64",
     F16 = "f16",
+    BF16 = "bf16",
     I8 = "i8",
     B1 = "b1"
 }

package/javascript/dist/cjs/usearch.js

@@ -74,6 +74,7 @@ var ScalarKind;
     ScalarKind["F32"] = "f32";
     ScalarKind["F64"] = "f64";
     ScalarKind["F16"] = "f16";
+    ScalarKind["BF16"] = "bf16";
     ScalarKind["I8"] = "i8";
     ScalarKind["B1"] = "b1";
 })(ScalarKind || (exports.ScalarKind = ScalarKind = {}));

package/javascript/dist/esm/usearch.d.ts

@@ -29,6 +29,7 @@ export declare enum ScalarKind {
     F32 = "f32",
     F64 = "f64",
     F16 = "f16",
+    BF16 = "bf16",
     I8 = "i8",
     B1 = "b1"
 }

package/javascript/dist/esm/usearch.js

@@ -33,6 +33,7 @@ export var ScalarKind;
     ScalarKind["F32"] = "f32";
     ScalarKind["F64"] = "f64";
     ScalarKind["F16"] = "f16";
+    ScalarKind["BF16"] = "bf16";
     ScalarKind["I8"] = "i8";
     ScalarKind["B1"] = "b1";
 })(ScalarKind || (ScalarKind = {}));

package/javascript/usearch.ts

@@ -70,6 +70,7 @@ export enum ScalarKind {
   F32 = "f32",
   F64 = "f64",
   F16 = "f16",
+  BF16 = "bf16",
   I8 = "i8",
   B1 = "b1",
 }
@@ -610,12 +611,12 @@ function exactSearch(
 }
 
 const usearch = {
-    Index,
-    MetricKind,
-    ScalarKind,
-    Matches,
-    BatchMatches,
-    exactSearch,
+  Index,
+  MetricKind,
+  ScalarKind,
+  Matches,
+  BatchMatches,
+  exactSearch,
 };
 export default usearch;
 

package/package.json

@@ -1,6 +1,6 @@
 {
     "name": "usearch",
-    "version": "2.13.4",
+    "version": "2.14.0",
     "description": "Smaller & Faster Single-File Vector Search Engine from Unum",
     "author": "Ash Vardanian (https://ashvardanian.com/)",
     "license": "Apache 2.0",