Fix: SQLite cosine function + tests

Author: ashvardanian

.vscode/settings.json

@@ -104,7 +104,8 @@
     "ranges": "cpp",
     "span": "cpp",
     "format": "cpp",
-    "charconv": "cpp"
+    "charconv": "cpp",
+    "strstream": "cpp"
   },
   "cSpell.words": [
     "allclose",
@@ -153,7 +154,9 @@
     "SLOC",
     "Sonatype",
     "sorensen",
+    "sqeuclidean",
     "Struct",
+    "swar",
     "tanimoto",
     "tqdm",
     "uninitialize",

python/lib_sqlite.cpp

@@ -79,7 +79,7 @@ static void sqlite_dense(sqlite3_context* context, int argc, sqlite3_value** arg
 
             // Parse the floating-point numbers
             std::from_chars_result result1 = std::from_chars(vec1, vec1 + bytes1, parsed1[i]);
-            std::from_chars_result result2 = std::from_chars(vec2, vec2 + bytes2, parsed1[i]);
+            std::from_chars_result result2 = std::from_chars(vec2, vec2 + bytes2, parsed2[i]);
             if (result1.ec != std::errc() || result2.ec != std::errc()) {
                 sqlite3_result_error(context, "Number can't be parsed", -1);
                 return;

python/scripts/test_sqlite.py

@@ -0,0 +1,80 @@
+import sqlite3
+import json
+import math
+import pytest
+import numpy as np
+
+import usearch
+from usearch.io import load_matrix, save_matrix
+from usearch.index import search
+from usearch.eval import random_vectors
+
+from usearch.index import Match, Matches, BatchMatches, Index, Indexes
+
+
+dimensions = [3, 97, 256]
+batch_sizes = [1, 77, 100]
+
+
+def test_sqlite_distances():
+    conn = sqlite3.connect(":memory:")
+    conn.enable_load_extension(True)
+    conn.load_extension(usearch.sqlite)
+
+    cursor = conn.cursor()
+
+    # Create a table with additional columns for f32 and f16 BLOBs
+    cursor.execute(
+        """
+        CREATE TABLE IF NOT EXISTS vector_table (
+            id INTEGER PRIMARY KEY,
+            vector_json JSON,
+            vector_f32 BLOB,
+            vector_f16 BLOB
+        )
+    """
+    )
+
+    # Generate and insert random vectors
+    num_vectors = 3  # Number of vectors to generate
+    dim = 4  # Dimension of each vector
+    vectors = []
+
+    for i in range(num_vectors):
+        # Generate a random 256-dimensional vector
+        vector = np.random.rand(dim)
+        vectors.append(vector)
+
+        # Convert the vector to f32 and f16
+        vector_f32 = np.float32(vector)
+        vector_f16 = np.float16(vector)
+
+        # Insert the vector into the database as JSON and as BLOBs
+        cursor.execute(
+            """
+            INSERT INTO vector_table (vector_json, vector_f32, vector_f16) VALUES (?, ?, ?)
+        """,
+            (json.dumps(vector.tolist()), vector_f32.tobytes(), vector_f16.tobytes()),
+        )
+
+    # Commit changes
+    conn.commit()
+
+    similarities = """
+    SELECT 
+        a.id AS id1,
+        b.id AS id2,
+        distance_cosine_f32(a.vector_json, b.vector_json) AS cosine_similarity_json,
+        distance_cosine_f32(a.vector_f32, b.vector_f32) AS cosine_similarity_f32,
+        distance_cosine_f16(a.vector_f16, b.vector_f16) AS cosine_similarity_f16
+    FROM 
+        vector_table AS a,
+        vector_table AS b
+    WHERE 
+        a.id < b.id;
+    """
+    cursor.execute(similarities)
+
+    for a, b, similarity_json, similarity_f32, similarity_f16 in cursor.fetchall():
+        assert math.isclose(similarity_json, similarity_f32, abs_tol=0.1)
+        assert math.isclose(similarity_json, similarity_f16, abs_tol=0.1)

python/usearch/__init__.py

@@ -1,7 +1,13 @@
+import importlib
+
 from usearch.compiled import (
     VERSION_MAJOR,
     VERSION_MINOR,
     VERSION_PATCH,
 )
 
 __version__ = f"{VERSION_MAJOR}.{VERSION_MINOR}.{VERSION_PATCH}"
+
+# The same binary file (.so, .dll, or .dylib) that contains the pre-compiled
+# USearch code also contains the SQLite3 binding
+sqlite = importlib.util.find_spec("usearch.compiled").origin