chore: GenAI Client(evals) - Pin all RubricMetric properties to v1

agentplatform/_genai/_evals_metric_loaders.py

@@ -277,27 +277,27 @@ def __getattr__(
 
     @property
     def GENERAL_QUALITY(self) -> LazyLoadedPrebuiltMetric:
-        return self.__getattr__("GENERAL_QUALITY")
+        return self.__getattr__("GENERAL_QUALITY", version="v1")
 
     @property
     def TEXT_QUALITY(self) -> LazyLoadedPrebuiltMetric:
-        return self.__getattr__("TEXT_QUALITY")
+        return self.__getattr__("TEXT_QUALITY", version="v1")
 
     @property
     def INSTRUCTION_FOLLOWING(self) -> LazyLoadedPrebuiltMetric:
-        return self.__getattr__("INSTRUCTION_FOLLOWING")
+        return self.__getattr__("INSTRUCTION_FOLLOWING", version="v1")
 
     @property
     def SAFETY(self) -> LazyLoadedPrebuiltMetric:
-        return self.__getattr__("SAFETY")
+        return self.__getattr__("SAFETY", version="v1")
 
     @property
     def MULTI_TURN_GENERAL_QUALITY(self) -> LazyLoadedPrebuiltMetric:
-        return self.__getattr__("MULTI_TURN_GENERAL_QUALITY")
+        return self.__getattr__("MULTI_TURN_GENERAL_QUALITY", version="v1")
 
     @property
     def MULTI_TURN_TEXT_QUALITY(self) -> LazyLoadedPrebuiltMetric:
-        return self.__getattr__("MULTI_TURN_TEXT_QUALITY")
+        return self.__getattr__("MULTI_TURN_TEXT_QUALITY", version="v1")
 
     @property
     def MULTI_TURN_TOOL_USE_QUALITY(self) -> LazyLoadedPrebuiltMetric:
@@ -317,43 +317,43 @@ def FINAL_RESPONSE_MATCH(self) -> LazyLoadedPrebuiltMetric:
 
     @property
     def FINAL_RESPONSE_REFERENCE_FREE(self) -> LazyLoadedPrebuiltMetric:
-        return self.__getattr__("FINAL_RESPONSE_REFERENCE_FREE")
+        return self.__getattr__("FINAL_RESPONSE_REFERENCE_FREE", version="v1")
 
     @property
     def COHERENCE(self) -> LazyLoadedPrebuiltMetric:
-        return self.__getattr__("COHERENCE")
+        return self.__getattr__("COHERENCE", version="v1")
 
     @property
     def FLUENCY(self) -> LazyLoadedPrebuiltMetric:
-        return self.__getattr__("FLUENCY")
+        return self.__getattr__("FLUENCY", version="v1")
 
     @property
     def VERBOSITY(self) -> LazyLoadedPrebuiltMetric:
-        return self.__getattr__("VERBOSITY")
+        return self.__getattr__("VERBOSITY", version="v1")
 
     @property
     def SUMMARIZATION_QUALITY(self) -> LazyLoadedPrebuiltMetric:
-        return self.__getattr__("SUMMARIZATION_QUALITY")
+        return self.__getattr__("SUMMARIZATION_QUALITY", version="v1")
 
     @property
     def QUESTION_ANSWERING_QUALITY(self) -> LazyLoadedPrebuiltMetric:
-        return self.__getattr__("QUESTION_ANSWERING_QUALITY")
+        return self.__getattr__("QUESTION_ANSWERING_QUALITY", version="v1")
 
     @property
     def MULTI_TURN_CHAT_QUALITY(self) -> LazyLoadedPrebuiltMetric:
-        return self.__getattr__("MULTI_TURN_CHAT_QUALITY")
+        return self.__getattr__("MULTI_TURN_CHAT_QUALITY", version="v1")
 
     @property
     def MULTI_TURN_SAFETY(self) -> LazyLoadedPrebuiltMetric:
-        return self.__getattr__("MULTI_TURN_SAFETY")
+        return self.__getattr__("MULTI_TURN_SAFETY", version="v1")
 
     @property
     def FINAL_RESPONSE_QUALITY(self) -> LazyLoadedPrebuiltMetric:
-        return self.__getattr__("FINAL_RESPONSE_QUALITY")
+        return self.__getattr__("FINAL_RESPONSE_QUALITY", version="v1")
 
     @property
     def HALLUCINATION(self) -> LazyLoadedPrebuiltMetric:
-        return self.__getattr__("HALLUCINATION")
+        return self.__getattr__("HALLUCINATION", version="v1")
 
     @property
     def GROUNDING(self) -> LazyLoadedPrebuiltMetric:  # pylint: disable=invalid-name
@@ -374,15 +374,15 @@ def GROUNDEDNESS(self) -> LazyLoadedPrebuiltMetric:  # pylint: disable=invalid-n
 
     @property
     def TOOL_USE_QUALITY(self) -> LazyLoadedPrebuiltMetric:
-        return self.__getattr__("TOOL_USE_QUALITY")
+        return self.__getattr__("TOOL_USE_QUALITY", version="v1")
 
     @property
     def GECKO_TEXT2IMAGE(self) -> LazyLoadedPrebuiltMetric:
-        return self.__getattr__("GECKO_TEXT2IMAGE")
+        return self.__getattr__("GECKO_TEXT2IMAGE", version="v1")
 
     @property
     def GECKO_TEXT2VIDEO(self) -> LazyLoadedPrebuiltMetric:
-        return self.__getattr__("GECKO_TEXT2VIDEO")
+        return self.__getattr__("GECKO_TEXT2VIDEO", version="v1")
 
 
 PrebuiltMetric = PrebuiltMetricLoader()

tests/unit/agentplatform/genai/test_evals.py

@@ -123,6 +123,7 @@ def mock_api_client_fixture():
 
 @pytest.fixture
 def mock_eval_dependencies(mock_api_client_fixture):
+    _evals_metric_loaders.LazyLoadedPrebuiltMetric._cache.clear()
     # fmt: off
     with (
         mock.patch("google.cloud.storage.Client") as mock_storage_client,
@@ -6386,6 +6387,49 @@ def test_groundedness_resolve_returns_grounding_v1_metric(self):
         assert resolved.name == "grounding_v1"
 
 
+class TestPrebuiltMetricLoaderVersionPinning:
+    """Verifies explicit version pinning for all RubricMetric properties."""
+
+    @pytest.mark.parametrize(
+        "prop_name,expected_spec",
+        [
+            ("GENERAL_QUALITY", "general_quality_v1"),
+            ("TEXT_QUALITY", "text_quality_v1"),
+            ("INSTRUCTION_FOLLOWING", "instruction_following_v1"),
+            ("SAFETY", "safety_v1"),
+            ("MULTI_TURN_GENERAL_QUALITY", "multi_turn_general_quality_v1"),
+            ("MULTI_TURN_TEXT_QUALITY", "multi_turn_text_quality_v1"),
+            ("FINAL_RESPONSE_REFERENCE_FREE", "final_response_reference_free_v1"),
+            ("FINAL_RESPONSE_QUALITY", "final_response_quality_v1"),
+            ("HALLUCINATION", "hallucination_v1"),
+            ("TOOL_USE_QUALITY", "tool_use_quality_v1"),
+            ("GECKO_TEXT2IMAGE", "gecko_text2image_v1"),
+            ("GECKO_TEXT2VIDEO", "gecko_text2video_v1"),
+        ],
+    )
+    def test_predefined_property_pins_to_v1(self, prop_name, expected_spec):
+        lazy_metric = getattr(agentplatform_genai_types.RubricMetric, prop_name)
+        assert lazy_metric.version == "v1"
+        assert lazy_metric._get_api_metric_spec_name() == expected_spec
+
+    @pytest.mark.parametrize(
+        "prop_name",
+        [
+            "COHERENCE",
+            "FLUENCY",
+            "VERBOSITY",
+            "SUMMARIZATION_QUALITY",
+            "QUESTION_ANSWERING_QUALITY",
+            "MULTI_TURN_CHAT_QUALITY",
+            "MULTI_TURN_SAFETY",
+        ],
+    )
+    def test_gcs_backed_property_pins_to_v1(self, prop_name):
+        lazy_metric = getattr(agentplatform_genai_types.RubricMetric, prop_name)
+        assert lazy_metric.version == "v1"
+        assert lazy_metric._get_api_metric_spec_name() is None
+
+
 class TestMergeResponseDatasets:
     """Unit tests for the merge_response_datasets_into_canonical_format function."""
 

tests/unit/vertexai/test_evaluation.py

@@ -1882,9 +1882,9 @@ def test_runnable_trajectory_eval_with_runnable_inference(self, api_transport):
                 "trajectory_exact_match/score",
             ]
         )
-        assert list(
+        assert sorted(
             test_result.metrics_table["trajectory_exact_match/score"].to_list()
-        ) == [1.0, 0.0]
+        ) == [0.0, 1.0]
 
     @pytest.mark.parametrize("api_transport", ["grpc", "rest"])
     def test_pointwise_autorater_request_config_enabled(self, api_transport):