diff --git a/src/vector_of_kll.cpp b/src/vector_of_kll.cpp
index 70139152..0f4fd64a 100644
--- a/src/vector_of_kll.cpp
+++ b/src/vector_of_kll.cpp
@@ -65,8 +65,10 @@ class vector_of_kll_sketches {
     void update(nb::ndarray<T>& items, char order);
     void merge(const vector_of_kll_sketches<T>& other);
 
+    // The array alternative is typed so that nanobind converts other dtypes
+    // (e.g. float32 ranks, int64 indices) instead of reinterpreting the buffer.
     template<typename V>
-    using ArrInputType = std::variant<nb::ndarray<>, nb::list, V>;
+    using ArrInputType = std::variant<nb::ndarray<V, nb::ndim<1>>, nb::list, V>;
 
     // returns a single sketch combining all data in the array
     kll_sketch<T, C> collapse(ArrInputType<int>& isk) const;
@@ -194,8 +196,8 @@ auto vector_of_kll_sketches<T, C>::make_ndarray(size_t rows, size_t cols) const
 template<typename T, typename C>
 template<typename TT>
 auto vector_of_kll_sketches<T, C>::input_to_vec(ArrInputType<TT>& input) const -> Array1D<TT> {
-  if (std::holds_alternative<nb::ndarray<>>(input)) {
-    nb::ndarray<> arr = std::get<nb::ndarray<>>(input);
+  if (std::holds_alternative<nb::ndarray<TT, nb::ndim<1>>>(input)) {
+    nb::ndarray<TT, nb::ndim<1>> arr = std::get<nb::ndarray<TT, nb::ndim<1>>>(input);
     return Array1D<TT>(arr);
   } else if (std::holds_alternative<nb::list>(input)) {
     nb::list list = std::get<nb::list>(input);
diff --git a/tests/vector_of_kll_test.py b/tests/vector_of_kll_test.py
index 51b2f0b2..adaae849 100644
--- a/tests/vector_of_kll_test.py
+++ b/tests/vector_of_kll_test.py
@@ -97,6 +97,27 @@ def test_kll_ints_sketches(self):
       kll = vector_of_kll_ints_sketches(k, d)
       self.assertTrue(np.all(kll.is_empty()))
 
+    def test_kll_input_dtypes(self):
+      # array inputs must give the same results as lists, whatever their dtype (issue #63)
+      kll = vector_of_kll_floats_sketches(200, 2)
+      kll.update(np.column_stack([np.arange(1, 101), np.arange(101, 201)]).astype(np.float32))
+
+      ranks = [0.0625, 0.9375] # exactly representable in float32
+      expected = kll.get_quantiles(ranks)
+      for dtype in (np.float32, np.float64):
+        np.testing.assert_array_equal(kll.get_quantiles(np.array(ranks, dtype=dtype)), expected)
+
+      values = [25.5, 150.5]
+      for fn in (kll.get_ranks, kll.get_pmf, kll.get_cdf):
+        expected = fn(values)
+        for dtype in (np.float32, np.float64):
+          np.testing.assert_array_equal(fn(np.array(values, dtype=dtype)), expected)
+
+      # sketch indices, including numpy's default int64
+      expected = kll.get_quantiles(0.5, [1, 0])
+      for dtype in (np.int32, np.int64):
+        np.testing.assert_array_equal(kll.get_quantiles(0.5, np.array([1, 0], dtype=dtype)), expected)
+
     def test_kll_2Dupdates(self):
       # 1D case tested in the first example
       # 2D case will follow same idea, but focusing on update()