Fixed handling of duplicates for SciPy sparse arrays - fixes #162

Author: ankane

pgvector/sparsevec.py

@@ -105,6 +105,11 @@ def _from_dict(self, d: dict[int, float], dim: int) -> None:
     def _from_sparse(self, arr: sparray | spmatrix, /) -> None:
         value: coo_array | coo_matrix = arr.tocoo(copy=False)  # type: ignore
 
+        # has_canonical_format added in scipy 1.12+
+        if not hasattr(value, 'has_canonical_format') or not value.has_canonical_format:
+            value = value.copy()
+            value.sum_duplicates()
+
         shape = cast(tuple[int, ...], value.shape)
         if len(shape) == 1:
             self._dim = shape[0]

tests/test_sparse_vector.py

@@ -60,6 +60,16 @@ def test_coo_array(self) -> None:
         assert vec.indices() == [0, 2, 4]
         assert isinstance(vec.values()[0], float)
 
+    def test_coo_array_duplicates(self) -> None:
+        if np is None or sparse is None:
+            pytest.skip('NumPy and SciPy required')
+
+        arr = sparse.coo_array(([1, 2], ([1, 1],)), shape=(3,))
+        assert arr.todense().tolist() == [0, 3, 0]
+        vec = SparseVector(arr)
+        assert vec.to_list() == [0, 3, 0]
+        assert vec.indices() == [1]
+
     def test_coo_array_dimensions(self) -> None:
         if np is None or sparse is None:
             pytest.skip('NumPy and SciPy required')