chore: de-flake sem_cluster test by relaxing the passing requirement (#1207)

Author: sycai

tests/system/large/operations/test_semantics.py

@@ -165,25 +165,37 @@ def test_agg_invalid_cluster_column_raise_error(gemini_flash_model, cluster_colu
     [
         pytest.param(1, id="one", marks=pytest.mark.xfail(raises=ValueError)),
         pytest.param(2, id="two"),
-        pytest.param(4, id="four"),
     ],
 )
 def test_cluster_by(session, text_embedding_generator, n_clusters):
     bigframes.options.experiments.semantic_operators = True
     df = dataframe.DataFrame(
-        ({"Product": ["Smartphone", "Laptop", "Coffee Maker", "T-shirt", "Jeans"]}),
+        (
+            {
+                "Item": [
+                    "Orange",
+                    "Cantaloupe",
+                    "Watermelon",
+                    "Chicken",
+                    "Duck",
+                    "Hen",
+                    "Rooster",
+                ]
+            }
+        ),
         session=session,
     )
     output_column = "cluster id"
     result = df.semantics.cluster_by(
-        "Product",
+        "Item",
         output_column,
         text_embedding_generator,
         n_clusters=n_clusters,
     )
 
     assert output_column in result
-    assert len(result[output_column].unique()) == n_clusters
+    # In rare cases, it's possible to have fewer than K clusters due to randomness.
+    assert len(result[output_column].unique()) <= n_clusters
 
 
 def test_cluster_by_invalid_column(session, text_embedding_generator):