feat: Add Safety evaluator metric

We add a new metric for evaluating safety of Agent's response to ADK Eval. We delegate the actual implementation to Vertex Gen AI Eval SDK, so using this metric will require GCP project.

As a part of this change, we created (refactored) a simple Facade for vertex gen ai eval sdk.

PiperOrigin-RevId: 778580406

Author: ankursharmas

src/google/adk/cli/cli_eval.py

@@ -41,6 +41,7 @@
 
 TOOL_TRAJECTORY_SCORE_KEY = "tool_trajectory_avg_score"
 RESPONSE_MATCH_SCORE_KEY = "response_match_score"
+SAFETY_V1_KEY = "safety_v1"
 # This evaluation is not very stable.
 # This is always optional unless explicitly specified.
 RESPONSE_EVALUATION_SCORE_KEY = "response_evaluation_score"
@@ -260,6 +261,7 @@ async def run_evals(
 def _get_evaluator(eval_metric: EvalMetric) -> Evaluator:
   try:
     from ..evaluation.response_evaluator import ResponseEvaluator
+    from ..evaluation.safety_evaluator import SafetyEvaluatorV1
     from ..evaluation.trajectory_evaluator import TrajectoryEvaluator
   except ModuleNotFoundError as e:
     raise ModuleNotFoundError(MISSING_EVAL_DEPENDENCIES_MESSAGE) from e
@@ -272,5 +274,7 @@ def _get_evaluator(eval_metric: EvalMetric) -> Evaluator:
     return ResponseEvaluator(
         threshold=eval_metric.threshold, metric_name=eval_metric.metric_name
     )
+  elif eval_metric.metric_name == SAFETY_V1_KEY:
+    return SafetyEvaluatorV1(eval_metric)
 
   raise ValueError(f"Unsupported eval metric: {eval_metric}")

src/google/adk/evaluation/agent_evaluator.py

@@ -30,6 +30,7 @@
 
 from .constants import MISSING_EVAL_DEPENDENCIES_MESSAGE
 from .eval_case import IntermediateData
+from .eval_metrics import EvalMetric
 from .eval_set import EvalSet
 from .evaluator import EvalStatus
 from .evaluator import EvaluationResult
@@ -46,11 +47,13 @@
 # This is always optional unless explicitly specified.
 RESPONSE_EVALUATION_SCORE_KEY = "response_evaluation_score"
 RESPONSE_MATCH_SCORE_KEY = "response_match_score"
+SAFETY_V1_KEY = "safety_v1"
 
 ALLOWED_CRITERIA = [
     TOOL_TRAJECTORY_SCORE_KEY,
     RESPONSE_EVALUATION_SCORE_KEY,
     RESPONSE_MATCH_SCORE_KEY,
+    SAFETY_V1_KEY,
 ]
 
 
@@ -387,6 +390,7 @@ def _validate_input(eval_dataset, criteria):
   def _get_metric_evaluator(metric_name: str, threshold: float) -> Evaluator:
     try:
       from .response_evaluator import ResponseEvaluator
+      from .safety_evaluator import SafetyEvaluatorV1
       from .trajectory_evaluator import TrajectoryEvaluator
     except ModuleNotFoundError as e:
       raise ModuleNotFoundError(MISSING_EVAL_DEPENDENCIES_MESSAGE) from e
@@ -397,6 +401,10 @@ def _get_metric_evaluator(metric_name: str, threshold: float) -> Evaluator:
         or metric_name == RESPONSE_EVALUATION_SCORE_KEY
     ):
       return ResponseEvaluator(threshold=threshold, metric_name=metric_name)
+    elif metric_name == SAFETY_V1_KEY:
+      return SafetyEvaluatorV1(
+          eval_metric=EvalMetric(threshold=threshold, metric_name=metric_name)
+      )
 
     raise ValueError(f"Unsupported eval metric: {metric_name}")
 

src/google/adk/evaluation/response_evaluator.py

@@ -14,26 +14,34 @@
 
 from __future__ import annotations
 
-import os
 from typing import Optional
 
-from google.genai import types as genai_types
-import pandas as pd
 from typing_extensions import override
-from vertexai import Client as VertexAiClient
 from vertexai import types as vertexai_types
 
 from .eval_case import Invocation
 from .eval_metrics import EvalMetric
-from .evaluator import EvalStatus
 from .evaluator import EvaluationResult
 from .evaluator import Evaluator
-from .evaluator import PerInvocationResult
 from .final_response_match_v1 import RougeEvaluator
+from .vertex_ai_eval_facade import _VertexAiEvalFacade
 
 
 class ResponseEvaluator(Evaluator):
-  """Runs response evaluation for agents."""
+  """Evaluates Agent's responses.
+
+  This class supports two metrics:
+  1) response_evaluation_score
+  This metric evaluates how coherent agent's resposne was.
+
+  Value range of this metric is [1,5], with values closer to 5 more desirable.
+
+  2) response_match_score:
+  This metric evaluates if agent's final response matches a golden/expected
+  final response.
+
+  Value range for this metric is [0,1], with values closer to 1 more desirable.
+  """
 
   def __init__(
       self,
@@ -77,80 +85,6 @@ def evaluate_invocations(
           actual_invocations, expected_invocations
       )
 
-    total_score = 0.0
-    num_invocations = 0
-    per_invocation_results = []
-    for actual, expected in zip(actual_invocations, expected_invocations):
-      prompt = self._get_text(expected.user_content)
-      reference = self._get_text(expected.final_response)
-      response = self._get_text(actual.final_response)
-
-      eval_case = {
-          "prompt": prompt,
-          "reference": reference,
-          "response": response,
-      }
-
-      eval_case_result = ResponseEvaluator._perform_eval(
-          pd.DataFrame([eval_case]), [self._metric_name]
-      )
-      score = self._get_score(eval_case_result)
-      per_invocation_results.append(
-          PerInvocationResult(
-              actual_invocation=actual,
-              expected_invocation=expected,
-              score=score,
-              eval_status=self._get_eval_status(score),
-          )
-      )
-
-      if score:
-        total_score += score
-        num_invocations += 1
-
-    if per_invocation_results:
-      overall_score = (
-          total_score / num_invocations if num_invocations > 0 else None
-      )
-      return EvaluationResult(
-          overall_score=overall_score,
-          overall_eval_status=self._get_eval_status(overall_score),
-          per_invocation_results=per_invocation_results,
-      )
-
-    return EvaluationResult()
-
-  def _get_text(self, content: Optional[genai_types.Content]) -> str:
-    if content and content.parts:
-      return "\n".join([p.text for p in content.parts if p.text])
-
-    return ""
-
-  def _get_score(self, eval_result) -> Optional[float]:
-    if eval_result and eval_result.summary_metrics:
-      return eval_result.summary_metrics[0].mean_score
-
-    return None
-
-  def _get_eval_status(self, score: Optional[float]):
-    if score:
-      return (
-          EvalStatus.PASSED if score >= self._threshold else EvalStatus.FAILED
-      )
-
-    return EvalStatus.NOT_EVALUATED
-
-  @staticmethod
-  def _perform_eval(dataset, metrics):
-    """This method hides away the call to external service.
-
-    Primarily helps with unit testing.
-    """
-    project_id = str(os.environ.get("GOOGLE_CLOUD_PROJECT"))
-    location = os.environ.get("GOOGLE_CLOUD_REGION")
-    client = VertexAiClient(project=project_id, location=location)
-
-    return client.evals.evaluate(
-        dataset=vertexai_types.EvaluationDataset(eval_dataset_df=dataset),
-        metrics=metrics,
-    )
+    return _VertexAiEvalFacade(
+        threshold=self._threshold, metric_name=self._metric_name
+    ).evaluate_invocations(actual_invocations, expected_invocations)

src/google/adk/evaluation/safety_evaluator.py

@@ -0,0 +1,54 @@
+# Copyright 2025 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import annotations
+
+from typing_extensions import override
+from vertexai import types as vertexai_types
+
+from .eval_case import Invocation
+from .eval_metrics import EvalMetric
+from .evaluator import EvaluationResult
+from .evaluator import Evaluator
+from .vertex_ai_eval_facade import _VertexAiEvalFacade
+
+
+class SafetyEvaluatorV1(Evaluator):
+  """Evaluates safety (harmlessness) of an Agent's Response.
+
+  The class delegates the responsibility to Vertex Gen AI Eval SDK. The V1
+  suffix in the class name is added to convey that there could be other versions
+  of the safety metric as well, and those metrics could use a different strategy
+  to evaluate safety.
+
+  Using this class requires a GCP project. Please set GOOGLE_CLOUD_PROJECT and
+  GOOGLE_CLOUD_LOCATION in your .env file.
+
+  Value range of the metric is [0, 1], with values closer to 1 to be more
+  desirable (safe).
+  """
+
+  def __init__(self, eval_metric: EvalMetric):
+    self._eval_metric = eval_metric
+
+  @override
+  def evaluate_invocations(
+      self,
+      actual_invocations: list[Invocation],
+      expected_invocations: list[Invocation],
+  ) -> EvaluationResult:
+    return _VertexAiEvalFacade(
+        threshold=self._eval_metric.threshold,
+        metric_name=vertexai_types.PrebuiltMetric.SAFETY,
+    ).evaluate_invocations(actual_invocations, expected_invocations)

src/google/adk/evaluation/vertex_ai_eval_facade.py

@@ -0,0 +1,147 @@
+# Copyright 2025 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import annotations
+
+import os
+from typing import Optional
+
+from google.genai import types as genai_types
+import pandas as pd
+from typing_extensions import override
+from vertexai import Client as VertexAiClient
+from vertexai import types as vertexai_types
+
+from .eval_case import Invocation
+from .evaluator import EvalStatus
+from .evaluator import EvaluationResult
+from .evaluator import Evaluator
+from .evaluator import PerInvocationResult
+
+_ERROR_MESSAGE_SUFFIX = """
+You should specify both project id and location. This metric uses Vertex Gen AI
+Eval SDK, and it requires google cloud credentials.
+
+If using an .env file add the values there, or explicitly set in the code using
+the template below:
+
+os.environ['GOOGLE_CLOUD_LOCATION'] = <LOCATION>
+os.environ['GOOGLE_CLOUD_PROJECT'] = <PROJECT ID>
+"""
+
+
+class _VertexAiEvalFacade(Evaluator):
+  """Simple facade for Vertex Gen AI Eval SDK.
+
+  Vertex Gen AI Eval SDK exposes quite a few metrics that are valuable for
+  agentic evals. This class helps us to access those metrics.
+
+  Using this class requires a GCP project. Please set GOOGLE_CLOUD_PROJECT and
+  GOOGLE_CLOUD_LOCATION in your .env file.
+  """
+
+  def __init__(
+      self, threshold: float, metric_name: vertexai_types.PrebuiltMetric
+  ):
+    self._threshold = threshold
+    self._metric_name = metric_name
+
+  @override
+  def evaluate_invocations(
+      self,
+      actual_invocations: list[Invocation],
+      expected_invocations: list[Invocation],
+  ) -> EvaluationResult:
+    total_score = 0.0
+    num_invocations = 0
+    per_invocation_results = []
+    for actual, expected in zip(actual_invocations, expected_invocations):
+      prompt = self._get_text(expected.user_content)
+      reference = self._get_text(expected.final_response)
+      response = self._get_text(actual.final_response)
+      eval_case = {
+          "prompt": prompt,
+          "reference": reference,
+          "response": response,
+      }
+
+      eval_case_result = _VertexAiEvalFacade._perform_eval(
+          dataset=pd.DataFrame([eval_case]), metrics=[self._metric_name]
+      )
+      score = self._get_score(eval_case_result)
+      per_invocation_results.append(
+          PerInvocationResult(
+              actual_invocation=actual,
+              expected_invocation=expected,
+              score=score,
+              eval_status=self._get_eval_status(score),
+          )
+      )
+
+      if score:
+        total_score += score
+        num_invocations += 1
+
+    if per_invocation_results:
+      overall_score = (
+          total_score / num_invocations if num_invocations > 0 else None
+      )
+      return EvaluationResult(
+          overall_score=overall_score,
+          overall_eval_status=self._get_eval_status(overall_score),
+          per_invocation_results=per_invocation_results,
+      )
+
+    return EvaluationResult()
+
+  def _get_text(self, content: Optional[genai_types.Content]) -> str:
+    if content and content.parts:
+      return "\n".join([p.text for p in content.parts if p.text])
+
+    return ""
+
+  def _get_score(self, eval_result) -> Optional[float]:
+    if eval_result and eval_result.summary_metrics:
+      return eval_result.summary_metrics[0].mean_score
+
+    return None
+
+  def _get_eval_status(self, score: Optional[float]):
+    if score:
+      return (
+          EvalStatus.PASSED if score >= self._threshold else EvalStatus.FAILED
+      )
+
+    return EvalStatus.NOT_EVALUATED
+
+  @staticmethod
+  def _perform_eval(dataset, metrics):
+    """This method hides away the call to external service.
+
+    Primarily helps with unit testing.
+    """
+    project_id = os.environ.get("GOOGLE_CLOUD_PROJECT", None)
+    location = os.environ.get("GOOGLE_CLOUD_LOCATION", None)
+
+    if not project_id:
+      raise ValueError("Missing project id." + _ERROR_MESSAGE_SUFFIX)
+    if not location:
+      raise ValueError("Missing location." + _ERROR_MESSAGE_SUFFIX)
+
+    client = VertexAiClient(project=project_id, location=location)
+
+    return client.evals.evaluate(
+        dataset=vertexai_types.EvaluationDataset(eval_dataset_df=dataset),
+        metrics=metrics,
+    )