usearch 0.17.1 → 0.18.1

package/.github/workflows/prerelease.yml

@@ -42,13 +42,13 @@ jobs:
         run: pytest python/scripts/test.py
 
 
-  test_python_36:
-    name: Test Python 3.6
+  test_python_37:
+    name: Test Python 3.7
     runs-on: ${{ matrix.os }}
     strategy:
       matrix:
         os: [ubuntu-20.04]
-        python-version: ["3.6"]
+        python-version: ["3.7"]
 
     steps:
       - uses: actions/checkout@v3
@@ -62,7 +62,7 @@ jobs:
       - name: Install dependencies
         run: |
           python -m pip install --upgrade pip
-          pip install pytest numpy dataclasses
+          pip install pytest numpy
 
       - name: Build locally
         run: python -m pip install .

package/.github/workflows/release.yml

@@ -125,6 +125,8 @@ jobs:
 
       - name: Build wheels
         run: python -m cibuildwheel
+        env:
+          CIBW_SKIP: cp36-*
 
       - uses: actions/upload-artifact@v3
         with:

package/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "usearch"
-version = "0.17.1"
+version = "0.18.1"
 authors = ["Ashot Vardanian <1983160+ashvardanian@users.noreply.github.com>"]
 description = "Smaller & Faster Single-File Vector Search Engine from Unum"
 edition = "2021"

package/README.md

@@ -276,7 +276,7 @@ The difference would be in the shape of the tensors.
 
 ```py
 n = 100
-labels = np.array(range(n), dtype=np.longlong)
+labels = np.arange(n)
 vectors = np.random.uniform(0, 0.3, (n, index.ndim)).astype(np.float32)
 
 index.add(labels, vectors, threads=..., copy=...)
@@ -584,7 +584,7 @@ fingerprints = np.vstack([encoder.GetFingerprint(x) for x in molecules])
 fingerprints = np.packbits(fingerprints, axis=1)
 
 index = Index(ndim=2048, metric=MetricKind.BitwiseTanimoto)
-labels = np.array(len(molecules), dtype=np.longlong)
+labels = np.arange(len(molecules))
 
 index.add(labels, fingerprints)
 matches = index.search(fingerprints, 10)

package/conanfile.py

@@ -4,7 +4,7 @@ from conans import ConanFile
 class USearchConan(ConanFile):
 
     name = 'USearch'
-    version = '0.17.1'
+    version = '0.18.1'
     license = 'Apache 2.0'
     url = 'https://github.com/unum-cloud/usearch'
     description = 'Smaller & Faster Single-File Vector Search Engine from Unum'

package/package.json

@@ -1,6 +1,6 @@
 {
     "name": "usearch",
-    "version": "0.17.1",
+    "version": "0.18.1",
     "description": "Smaller & Faster Single-File Vector Search Engine from Unum",
     "author": "Ashot Vardanian",
     "license": "Apache 2.0",

package/python/lib.cpp

@@ -26,7 +26,7 @@ using namespace unum;
 namespace py = pybind11;
 using py_shape_t = py::array::ShapeContainer;
 
-using label_t = Py_ssize_t;
+using label_t = std::uint32_t;
 using distance_t = punned_distance_t;
 using metric_t = punned_metric_t;
 using id_t = std::uint32_t;
@@ -132,11 +132,11 @@ static void add_one_to_index(punned_index_py_t& index, label_t label, py::buffer
     config.store_vector = copy;
 
     // https://docs.python.org/3/library/struct.html#format-characters
-    if (vector_info.format == "B")
+    if (vector_info.format == "B" || vector_info.format == "u1" || vector_info.format == "|u1")
         index.add(label, reinterpret_cast<b1x8_t const*>(vector_data), config).error.raise();
-    else if (vector_info.format == "c" || vector_info.format == "b")
+    else if (vector_info.format == "b" || vector_info.format == "i1" || vector_info.format == "|i1")
         index.add(label, reinterpret_cast<f8_bits_t const*>(vector_data), config).error.raise();
-    else if (vector_info.format == "e")
+    else if (vector_info.format == "e" || vector_info.format == "f2" || vector_info.format == "<f2")
         index.add(label, reinterpret_cast<f16_t const*>(vector_data), config).error.raise();
     else if (vector_info.format == "f" || vector_info.format == "f4" || vector_info.format == "<f4")
         index.add(label, reinterpret_cast<float const*>(vector_data), config).error.raise();
@@ -156,7 +156,7 @@ static void add_many_to_index(                                       //
     py::buffer_info labels_info = labels.request();
     py::buffer_info vectors_info = vectors.request();
 
-    if (labels_info.format != py::format_descriptor<label_t>::format())
+    if (labels_info.itemsize != sizeof(label_t))
         throw std::invalid_argument("Incompatible label type!");
 
     if (labels_info.ndim != 1)
@@ -181,7 +181,7 @@ static void add_many_to_index(                                       //
     char const* labels_data = reinterpret_cast<char const*>(labels_info.ptr);
 
     // https://docs.python.org/3/library/struct.html#format-characters
-    if (vectors_info.format == "B")
+    if (vectors_info.format == "B" || vectors_info.format == "u1" || vectors_info.format == "|u1")
         executor_default_t{threads}.execute_bulk(vectors_count, [&](std::size_t thread_idx, std::size_t task_idx) {
             add_config_t config;
             config.store_vector = copy;
@@ -190,7 +190,7 @@ static void add_many_to_index(                                       //
             b1x8_t const* vector = reinterpret_cast<b1x8_t const*>(vectors_data + task_idx * vectors_info.strides[0]);
             index.add(label, vector, config).error.raise();
         });
-    else if (vectors_info.format == "c" || vectors_info.format == "b")
+    else if (vectors_info.format == "b" || vectors_info.format == "i1" || vectors_info.format == "|i1")
         executor_default_t{threads}.execute_bulk(vectors_count, [&](std::size_t thread_idx, std::size_t task_idx) {
             add_config_t config;
             config.store_vector = copy;
@@ -200,7 +200,7 @@ static void add_many_to_index(                                       //
                 reinterpret_cast<f8_bits_t const*>(vectors_data + task_idx * vectors_info.strides[0]);
             index.add(label, vector, config).error.raise();
         });
-    else if (vectors_info.format == "e")
+    else if (vectors_info.format == "e" || vectors_info.format == "f2" || vectors_info.format == "<f2")
         executor_default_t{threads}.execute_bulk(vectors_count, [&](std::size_t thread_idx, std::size_t task_idx) {
             add_config_t config;
             config.store_vector = copy;
@@ -249,15 +249,15 @@ static py::tuple search_one_in_index(punned_index_py_t& index, py::buffer vector
     config.exact = exact;
 
     // https://docs.python.org/3/library/struct.html#format-characters
-    if (vector_info.format == "B") {
+    if (vector_info.format == "B" || vector_info.format == "u1" || vector_info.format == "|u1") {
         punned_search_result_t result = index.search(reinterpret_cast<b1x8_t const*>(vector_data), wanted, config);
         result.error.raise();
         count = result.dump_to(&labels_py1d(0), &distances_py1d(0));
-    } else if (vector_info.format == "c" || vector_info.format == "b") {
+    } else if (vector_info.format == "b" || vector_info.format == "i1" || vector_info.format == "|i1") {
         punned_search_result_t result = index.search(reinterpret_cast<f8_bits_t const*>(vector_data), wanted, config);
         result.error.raise();
         count = result.dump_to(&labels_py1d(0), &distances_py1d(0));
-    } else if (vector_info.format == "e") {
+    } else if (vector_info.format == "e" || vector_info.format == "f2" || vector_info.format == "<f2") {
         punned_search_result_t result = index.search(reinterpret_cast<f16_t const*>(vector_data), wanted, config);
         result.error.raise();
         count = result.dump_to(&labels_py1d(0), &distances_py1d(0));
@@ -320,7 +320,7 @@ static py::tuple search_many_in_index( //
     auto counts_py1d = counts_py.template mutable_unchecked<1>();
 
     // https://docs.python.org/3/library/struct.html#format-characters
-    if (vectors_info.format == "B")
+    if (vectors_info.format == "B" || vectors_info.format == "u1" || vectors_info.format == "|u1")
         executor_default_t{threads}.execute_bulk(vectors_count, [&](std::size_t thread_idx, std::size_t task_idx) {
             search_config_t config;
             config.thread = thread_idx;
@@ -331,7 +331,7 @@ static py::tuple search_many_in_index( //
             counts_py1d(task_idx) =
                 static_cast<Py_ssize_t>(result.dump_to(&labels_py2d(task_idx, 0), &distances_py2d(task_idx, 0)));
         });
-    else if (vectors_info.format == "c" || vectors_info.format == "b")
+    else if (vectors_info.format == "b" || vectors_info.format == "i1" || vectors_info.format == "|i1")
         executor_default_t{threads}.execute_bulk(vectors_count, [&](std::size_t thread_idx, std::size_t task_idx) {
             search_config_t config;
             config.thread = thread_idx;
@@ -342,7 +342,7 @@ static py::tuple search_many_in_index( //
             counts_py1d(task_idx) =
                 static_cast<Py_ssize_t>(result.dump_to(&labels_py2d(task_idx, 0), &distances_py2d(task_idx, 0)));
         });
-    else if (vectors_info.format == "e")
+    else if (vectors_info.format == "e" || vectors_info.format == "f2" || vectors_info.format == "<f2")
         executor_default_t{threads}.execute_bulk(vectors_count, [&](std::size_t thread_idx, std::size_t task_idx) {
             search_config_t config;
             config.thread = thread_idx;

package/python/usearch/eval.py

@@ -1,21 +1,13 @@
-from typing import Tuple, Any, Callable, Union, Optional, List, NamedTuple
+from __future__ import annotations
+from time import time_ns
+from typing import Tuple, Any, Callable, Union, Optional, List
+from dataclasses import dataclass, asdict
 from collections import defaultdict
 
 import numpy as np
 
 from usearch.io import load_matrix
-from usearch.index import Index, Matches, MetricKind, MetricKindBitwise
-
-
-# For compatibility with Python 3.6
-try:
-    from time import time_ns
-except ImportError:
-    from datetime import datetime
-
-    def time_ns():
-        now = datetime.now()
-        return int(now.timestamp() * 1e9)
+from usearch.index import Index, Matches, MetricKind, MetricKindBitwise, Label
 
 
 def random_vectors(
@@ -89,7 +81,8 @@ def measure_seconds(f: Callable) -> Tuple[float, Any]:
     return secs, result
 
 
-class Dataset(NamedTuple):
+@dataclass
+class Dataset:
 
     labels: np.ndarray
     vectors: np.ndarray
@@ -134,7 +127,7 @@ class Dataset(NamedTuple):
             count = min(
                 d.vectors.shape[0], count) if count is not None else d.vectors.shape[0]
             d.vectors = d.vectors[:count, :]
-            d.labels = np.arange(count, dtype=np.longlong)
+            d.labels = np.arange(count, dtype=Label)
 
             if queries is not None:
                 d.queries = load_matrix(queries)
@@ -157,13 +150,14 @@ class Dataset(NamedTuple):
 
             d.vectors = random_vectors(count=count, ndim=ndim)
             d.queries = d.vectors
-            d.labels = np.arange(count, dtype=np.longlong)
+            d.labels = np.arange(count, dtype=Label)
             d.neighbors = np.reshape(d.labels, (count, 1))
 
         return d
 
 
-class TaskResult(NamedTuple):
+@dataclass
+class TaskResult:
 
     add_operations: Optional[int] = None
     add_per_second: Optional[float] = None
@@ -190,7 +184,7 @@ class TaskResult(NamedTuple):
     def search_seconds(self) -> float:
         return self.search_operations / self.search_per_second
 
-    def __add__(self, other):
+    def __add__(self, other: TaskResult):
         result = TaskResult()
         if self.add_operations and other.add_operations:
             result.add_operations = self.add_operations + other.add_operations
@@ -217,7 +211,8 @@ class TaskResult(NamedTuple):
         return result
 
 
-class AddTask(NamedTuple):
+@dataclass
+class AddTask:
 
     labels: np.ndarray
     vectors: np.ndarray
@@ -251,7 +246,7 @@ class AddTask(NamedTuple):
         self.labels = self.labels[new_order]
         self.vectors = self.vectors[new_order, :]
 
-    def slices(self, batch_size: int) -> list:
+    def slices(self, batch_size: int) -> List[AddTask]:
         """Splits this dataset into smaller chunks."""
 
         return [
@@ -260,7 +255,7 @@ class AddTask(NamedTuple):
                 vectors=self.vectors[start_row:start_row+batch_size, :],
             ) for start_row in range(0, self.count, batch_size)]
 
-    def clusters(self, number_of_clusters: int) -> list:
+    def clusters(self, number_of_clusters: int) -> List[AddTask]:
         """Splits this dataset into smaller chunks."""
 
         from sklearn.cluster import KMeans
@@ -280,7 +275,8 @@ class AddTask(NamedTuple):
             ) for rows in partitioning.values()]
 
 
-class SearchTask(NamedTuple):
+@dataclass
+class SearchTask:
 
     queries: np.ndarray
     neighbors: np.ndarray
@@ -295,7 +291,7 @@ class SearchTask(NamedTuple):
             recall_at_one=results.recall_first(self.neighbors[:, 0].flatten()),
         )
 
-    def slices(self, batch_size: int) -> list:
+    def slices(self, batch_size: int) -> List[SearchTask]:
         """Splits this dataset into smaller chunks."""
 
         return [
@@ -305,7 +301,8 @@ class SearchTask(NamedTuple):
             ) for start_row in range(0, self.queries.shape[0], batch_size)]
 
 
-class Evaluation(NamedTuple):
+@dataclass
+class Evaluation:
 
     tasks: List[Union[AddTask, SearchTask]]
     count: int
@@ -315,7 +312,7 @@ class Evaluation(NamedTuple):
     def for_dataset(
             dataset: Dataset,
             batch_size: int = 0,
-            clusters: int = 1):
+            clusters: int = 1) -> Evaluation:
 
         tasks = []
         add = AddTask(
@@ -356,5 +353,5 @@ class Evaluation(NamedTuple):
             index.clear()
         return {
             **index.specs,
-            **task_result._asdict(),
+            **asdict(task_result),
         }

package/python/usearch/index.py

@@ -29,7 +29,7 @@ MetricKindBitwise = (
 
 SparseIndex = _CompiledSetsIndex
 
-Label = np.longlong
+Label = np.uint32
 
 
 def _normalize_dtype(dtype) -> ScalarKind:
@@ -295,9 +295,9 @@ class Index:
             else:
                 labels = start_id
         if isinstance(labels, np.ndarray):
-            labels = labels.astype(np.longlong)
+            labels = labels.astype(Label)
         elif isinstance(labels, Iterable):
-            labels = np.array(labels, dtype=np.longlong)
+            labels = np.array(labels, dtype=Label)
 
         self._compiled.add(labels, vectors, copy=copy, threads=threads)
 

package/python/usearch/server.py

@@ -7,7 +7,7 @@ import numpy as np
 from typing import List
 
 from ucall.rich_posix import Server
-from usearch.index import Index, list_matches
+from usearch.index import Index, Matches, Label
 
 
 def _ascii_to_vector(string: str) -> np.ndarray:
@@ -58,26 +58,25 @@ def serve(
     @server
     def add_one(label: int, vector: np.ndarray):
         print('adding', label, vector)
-        labels = np.array([label], dtype=np.longlong)
+        labels = np.array([label], dtype=Label)
         vectors = vector.flatten().reshape(vector.shape[0], 1)
         index.add(labels, vectors)
 
     @server
     def add_many(labels: np.ndarray, vectors: np.ndarray):
-        labels = labels.astype(np.longlong)
         index.add(labels, vectors, threads=threads)
 
     @server
     def search_one(vector: np.ndarray, count: int) -> List[dict]:
         print('search', vector, count)
         vectors = vector.reshape(vector.shape[0], 1)
-        results = index.search(vectors, count)
-        return list_matches(results, 0)
+        results: Matches = index.search(vectors, count)
+        return results.to_list()
 
     @server
     def search_many(vectors: np.ndarray, count: int) -> List[List[dict]]:
-        results = index.search(vectors, count)
-        return [list_matches(results, i) for i in range(vectors.shape[0])]
+        results: Matches = index.search(vectors, count)
+        return results.to_list()
 
     @server
     def add_ascii(label: int, string: str):