diff --git a/implicit/nearest_neighbours.py b/implicit/nearest_neighbours.py
index a1f36c1..29e50c2 100644
--- a/implicit/nearest_neighbours.py
+++ b/implicit/nearest_neighbours.py
@@ -62,6 +62,7 @@ def recommend(
                 userid,
                 user_items=user_items,
                 N=N,
+                score_dtype=np.float64,
                 filter_already_liked_items=filter_already_liked_items,
                 filter_items=filter_items,
                 recalculate_user=recalculate_user,
@@ -115,7 +116,12 @@ def similar_items(
 
         if not np.isscalar(itemid):
             return _batch_call(
-                self.similar_items, itemid, N=N, filter_items=filter_items, items=items
+                self.similar_items,
+                itemid,
+                N=N,
+                score_dtype=np.float64,
+                filter_items=filter_items,
+                items=items,
             )
 
         if filter_items is not None and items is not None:
diff --git a/implicit/utils.py b/implicit/utils.py
index f61da18..6b7ac5a 100644
--- a/implicit/utils.py
+++ b/implicit/utils.py
@@ -103,11 +103,11 @@ def augment_inner_product_matrix(factors):
     return max_norm, np.append(factors, extra_dimension.reshape(norms.shape[0], 1), axis=1)
 
 
-def _batch_call(func, ids, *args, N=10, **kwargs):
+def _batch_call(func, ids, *args, N=10, id_dtype=np.int32, score_dtype=np.float32, **kwargs):
     # we're running in batch mode, just loop over each item and call the scalar version of the
     # function
-    output_ids = np.zeros((len(ids), N), dtype=np.int32)
-    output_scores = np.zeros((len(ids), N), dtype=np.float32)
+    output_ids = np.zeros((len(ids), N), dtype=id_dtype)
+    output_scores = np.zeros((len(ids), N), dtype=score_dtype)
 
     user_items = kwargs.pop("user_items") if "user_items" in kwargs else None
     item_users = kwargs.pop("item_users") if "item_users" in kwargs else None