diff --git a/api/python/cellxgene_census/src/cellxgene_census/experimental/_embedding_search.py b/api/python/cellxgene_census/src/cellxgene_census/experimental/_embedding_search.py
index de09e2060..3fc628e78 100644
--- a/api/python/cellxgene_census/src/cellxgene_census/experimental/_embedding_search.py
+++ b/api/python/cellxgene_census/src/cellxgene_census/experimental/_embedding_search.py
@@ -160,16 +160,17 @@ def predict_obs_metadata(
         # TODO: something more intelligent for numeric columns! also use distances, etc.
         max_joinid = neighbor_obs.index.max()
         out: dict[str, pd.Series[Any]] = {}
-        indices = np.broadcast_to(np.arange(neighbors.neighbor_ids.shape[0]), (10, neighbors.neighbor_ids.shape[0])).T
+        n_queries, n_neighbors = neighbors.neighbor_ids.shape
+        indices = np.broadcast_to(np.arange(n_queries), (n_neighbors, n_queries)).T
         g = sparse.csr_matrix(
             (
-                np.broadcast_to(1, neighbors.neighbor_ids.shape[0] * 10),
+                np.broadcast_to(1, n_queries * n_neighbors),
                 (
                     indices.flatten(),
                     neighbors.neighbor_ids.astype(np.int64).flatten(),
                 ),
             ),
-            shape=(neighbors.neighbor_ids.shape[0], max_joinid + 1),
+            shape=(n_queries, max_joinid + 1),
         )
         for col in column_names:
             col_categorical = neighbor_obs[col].astype("category")
diff --git a/api/python/cellxgene_census/tests/experimental/test_embeddings_search.py b/api/python/cellxgene_census/tests/experimental/test_embeddings_search.py
index a58a31628..9e1c3da0a 100644
--- a/api/python/cellxgene_census/tests/experimental/test_embeddings_search.py
+++ b/api/python/cellxgene_census/tests/experimental/test_embeddings_search.py
@@ -39,6 +39,26 @@ def test_embeddings_search(true_neighbors: dict[str, Any], query_result: Neighbo
     return
 
 
+@pytest.mark.experimental
+@pytest.mark.live_corpus
+@pytest.mark.parametrize("n_neighbors", [5, 7, 20])
+def test_embedding_search_n_neighbors(query_anndata: ad.AnnData, n_neighbors: int) -> None:
+    columns = ["cell_type"]
+    result = find_nearest_obs(
+        TRUE_NEAREST_NEIGHBORS_EMBEDDING_NAME,
+        TRUE_NEAREST_NEIGHBORS_ORGANISM,
+        TRUE_NEAREST_NEIGHBORS_CENSUS_VERSION,
+        query_anndata,
+        k=n_neighbors,
+        nprobe=25,
+    )
+
+    # Check that the correct number of neighbors is being returned
+    assert result.neighbor_ids.shape[1] == n_neighbors
+    # Check that this step works
+    _ = predict_obs_metadata(TRUE_NEAREST_NEIGHBORS_ORGANISM, TRUE_NEAREST_NEIGHBORS_CENSUS_VERSION, result, columns)
+
+
 @pytest.mark.experimental
 @pytest.mark.live_corpus
 def test_embeddings_search_errors(query_anndata: ad.AnnData) -> None: