From 0bd05df471a440159a44b5864be4740b0f1565f9 Mon Sep 17 00:00:00 2001 From: Ankur Sharma Date: Wed, 2 Jul 2025 11:29:41 -0700 Subject: [PATCH] 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 --- src/google/adk/cli/cli_eval.py | 4 + src/google/adk/evaluation/agent_evaluator.py | 8 + .../adk/evaluation/response_evaluator.py | 102 ++------ src/google/adk/evaluation/safety_evaluator.py | 54 +++++ .../adk/evaluation/vertex_ai_eval_facade.py | 147 ++++++++++++ .../hello_world_agent/test_config.json | 3 +- .../home_automation_agent/test_config.json | 3 +- .../evaluation/test_response_evaluator.py | 149 +----------- .../evaluation/test_safety_evaluator.py | 78 ++++++ .../evaluation/test_vertex_ai_eval_facade.py | 226 ++++++++++++++++++ 10 files changed, 544 insertions(+), 230 deletions(-) create mode 100644 src/google/adk/evaluation/safety_evaluator.py create mode 100644 src/google/adk/evaluation/vertex_ai_eval_facade.py create mode 100644 tests/unittests/evaluation/test_safety_evaluator.py create mode 100644 tests/unittests/evaluation/test_vertex_ai_eval_facade.py diff --git a/src/google/adk/cli/cli_eval.py b/src/google/adk/cli/cli_eval.py index 01b06135..d122c215 100644 --- a/src/google/adk/cli/cli_eval.py +++ b/src/google/adk/cli/cli_eval.py @@ -41,6 +41,7 @@ logger = logging.getLogger("google_adk." + __name__) 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}") diff --git a/src/google/adk/evaluation/agent_evaluator.py b/src/google/adk/evaluation/agent_evaluator.py index 486d01cf..27c35c66 100644 --- a/src/google/adk/evaluation/agent_evaluator.py +++ b/src/google/adk/evaluation/agent_evaluator.py @@ -30,6 +30,7 @@ from pydantic import ValidationError 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 @@ TOOL_TRAJECTORY_SCORE_KEY = "tool_trajectory_avg_score" # 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 @@ class AgentEvaluator: 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 @@ class AgentEvaluator: 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}") diff --git a/src/google/adk/evaluation/response_evaluator.py b/src/google/adk/evaluation/response_evaluator.py index 1a73df87..b38d5553 100644 --- a/src/google/adk/evaluation/response_evaluator.py +++ b/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 @@ class ResponseEvaluator(Evaluator): 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) diff --git a/src/google/adk/evaluation/safety_evaluator.py b/src/google/adk/evaluation/safety_evaluator.py new file mode 100644 index 00000000..6b9ad242 --- /dev/null +++ b/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) diff --git a/src/google/adk/evaluation/vertex_ai_eval_facade.py b/src/google/adk/evaluation/vertex_ai_eval_facade.py new file mode 100644 index 00000000..7adab34e --- /dev/null +++ b/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'] = +os.environ['GOOGLE_CLOUD_PROJECT'] = +""" + + +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, + ) diff --git a/tests/integration/fixture/hello_world_agent/test_config.json b/tests/integration/fixture/hello_world_agent/test_config.json index c7fba6a4..87393e02 100644 --- a/tests/integration/fixture/hello_world_agent/test_config.json +++ b/tests/integration/fixture/hello_world_agent/test_config.json @@ -1,6 +1,7 @@ { "criteria": { "tool_trajectory_avg_score": 1.0, - "response_match_score": 0.5 + "response_match_score": 0.5, + "safety_v1": 0.8 } } diff --git a/tests/integration/fixture/home_automation_agent/test_config.json b/tests/integration/fixture/home_automation_agent/test_config.json index 424c95de..56817a7e 100644 --- a/tests/integration/fixture/home_automation_agent/test_config.json +++ b/tests/integration/fixture/home_automation_agent/test_config.json @@ -1,5 +1,6 @@ { "criteria": { - "tool_trajectory_avg_score": 1.0 + "tool_trajectory_avg_score": 1.0, + "safety_v1": 0.8 } } diff --git a/tests/unittests/evaluation/test_response_evaluator.py b/tests/unittests/evaluation/test_response_evaluator.py index a1dc3aac..09946772 100644 --- a/tests/unittests/evaluation/test_response_evaluator.py +++ b/tests/unittests/evaluation/test_response_evaluator.py @@ -13,7 +13,6 @@ # limitations under the License. """Tests for the Response Evaluator.""" -import random from unittest.mock import patch from google.adk.evaluation.eval_case import Invocation @@ -25,7 +24,7 @@ from vertexai import types as vertexai_types @patch( - "google.adk.evaluation.response_evaluator.ResponseEvaluator._perform_eval" + "google.adk.evaluation.vertex_ai_eval_facade._VertexAiEvalFacade._perform_eval" ) class TestResponseEvaluator: """A class to help organize "patch" that are applicable to all tests.""" @@ -109,146 +108,8 @@ class TestResponseEvaluator: assert evaluation_result.overall_score == 0.9 assert evaluation_result.overall_eval_status == EvalStatus.PASSED mock_perform_eval.assert_called_once() - - def test_evaluate_invocations_coherence_metric_failed( - self, mock_perform_eval - ): - """Test evaluate_invocations function for Coherence metric.""" - actual_invocations = [ - Invocation( - user_content=genai_types.Content( - parts=[genai_types.Part(text="This is a test query.")] - ), - final_response=genai_types.Content( - parts=[ - genai_types.Part(text="This is a test candidate response.") - ] - ), - ) + _, mock_kwargs = mock_perform_eval.call_args + # Compare the names of the metrics. + assert [m.name for m in mock_kwargs["metrics"]] == [ + vertexai_types.PrebuiltMetric.COHERENCE.name ] - expected_invocations = [ - Invocation( - user_content=genai_types.Content( - parts=[genai_types.Part(text="This is a test query.")] - ), - final_response=genai_types.Content( - parts=[genai_types.Part(text="This is a test reference.")] - ), - ) - ] - evaluator = ResponseEvaluator( - threshold=0.8, metric_name="response_evaluation_score" - ) - # Mock the return value of _perform_eval - mock_perform_eval.return_value = vertexai_types.EvaluationResult( - summary_metrics=[vertexai_types.AggregatedMetricResult(mean_score=0.7)], - eval_case_results=[], - ) - - evaluation_result = evaluator.evaluate_invocations( - actual_invocations, expected_invocations - ) - - assert evaluation_result.overall_score == 0.7 - assert evaluation_result.overall_eval_status == EvalStatus.FAILED - mock_perform_eval.assert_called_once() - - def test_evaluate_invocations_coherence_metric_no_score( - self, mock_perform_eval - ): - """Test evaluate_invocations function for Coherence metric.""" - actual_invocations = [ - Invocation( - user_content=genai_types.Content( - parts=[genai_types.Part(text="This is a test query.")] - ), - final_response=genai_types.Content( - parts=[ - genai_types.Part(text="This is a test candidate response.") - ] - ), - ) - ] - expected_invocations = [ - Invocation( - user_content=genai_types.Content( - parts=[genai_types.Part(text="This is a test query.")] - ), - final_response=genai_types.Content( - parts=[genai_types.Part(text="This is a test reference.")] - ), - ) - ] - evaluator = ResponseEvaluator( - threshold=0.8, metric_name="response_evaluation_score" - ) - # Mock the return value of _perform_eval - mock_perform_eval.return_value = vertexai_types.EvaluationResult( - summary_metrics=[], - eval_case_results=[], - ) - - evaluation_result = evaluator.evaluate_invocations( - actual_invocations, expected_invocations - ) - - assert evaluation_result.overall_score is None - assert evaluation_result.overall_eval_status == EvalStatus.NOT_EVALUATED - mock_perform_eval.assert_called_once() - - def test_evaluate_invocations_coherence_metric_multiple_invocations( - self, mock_perform_eval - ): - """Test evaluate_invocations function for Coherence metric with multiple invocations.""" - num_invocations = 6 - actual_invocations = [] - expected_invocations = [] - mock_eval_results = [] - random.seed(61553) - scores = [random.random() for _ in range(num_invocations)] - - for i in range(num_invocations): - actual_invocations.append( - Invocation( - user_content=genai_types.Content( - parts=[genai_types.Part(text=f"Query {i+1}")] - ), - final_response=genai_types.Content( - parts=[genai_types.Part(text=f"Response {i+1}")] - ), - ) - ) - expected_invocations.append( - Invocation( - user_content=genai_types.Content( - parts=[genai_types.Part(text=f"Query {i+1}")] - ), - final_response=genai_types.Content( - parts=[genai_types.Part(text=f"Reference {i+1}")] - ), - ) - ) - mock_eval_results.append( - vertexai_types.EvaluationResult( - summary_metrics=[ - vertexai_types.AggregatedMetricResult(mean_score=scores[i]) - ], - eval_case_results=[], - ) - ) - - evaluator = ResponseEvaluator( - threshold=0.8, metric_name="response_evaluation_score" - ) - # Mock the return value of _perform_eval - mock_perform_eval.side_effect = mock_eval_results - - evaluation_result = evaluator.evaluate_invocations( - actual_invocations, expected_invocations - ) - - assert evaluation_result.overall_score == pytest.approx( - sum(scores) / num_invocations - ) - assert evaluation_result.overall_eval_status == EvalStatus.FAILED - assert mock_perform_eval.call_count == num_invocations diff --git a/tests/unittests/evaluation/test_safety_evaluator.py b/tests/unittests/evaluation/test_safety_evaluator.py new file mode 100644 index 00000000..077e3143 --- /dev/null +++ b/tests/unittests/evaluation/test_safety_evaluator.py @@ -0,0 +1,78 @@ +# 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. + +"""Tests for the Response Evaluator.""" +from unittest.mock import patch + +from google.adk.evaluation.eval_case import Invocation +from google.adk.evaluation.eval_metrics import EvalMetric +from google.adk.evaluation.evaluator import EvalStatus +from google.adk.evaluation.safety_evaluator import SafetyEvaluatorV1 +from google.genai import types as genai_types +from vertexai import types as vertexai_types + + +@patch( + "google.adk.evaluation.vertex_ai_eval_facade._VertexAiEvalFacade._perform_eval" +) +class TestSafetyEvaluatorV1: + """A class to help organize "patch" that are applicable to all tests.""" + + def test_evaluate_invocations_coherence_metric_passed( + self, mock_perform_eval + ): + """Test evaluate_invocations function for Coherence metric.""" + actual_invocations = [ + Invocation( + user_content=genai_types.Content( + parts=[genai_types.Part(text="This is a test query.")] + ), + final_response=genai_types.Content( + parts=[ + genai_types.Part(text="This is a test candidate response.") + ] + ), + ) + ] + expected_invocations = [ + Invocation( + user_content=genai_types.Content( + parts=[genai_types.Part(text="This is a test query.")] + ), + final_response=genai_types.Content( + parts=[genai_types.Part(text="This is a test reference.")] + ), + ) + ] + evaluator = SafetyEvaluatorV1( + eval_metric=EvalMetric(threshold=0.8, metric_name="safety") + ) + # Mock the return value of _perform_eval + mock_perform_eval.return_value = vertexai_types.EvaluationResult( + summary_metrics=[vertexai_types.AggregatedMetricResult(mean_score=0.9)], + eval_case_results=[], + ) + + evaluation_result = evaluator.evaluate_invocations( + actual_invocations, expected_invocations + ) + + assert evaluation_result.overall_score == 0.9 + assert evaluation_result.overall_eval_status == EvalStatus.PASSED + mock_perform_eval.assert_called_once() + _, mock_kwargs = mock_perform_eval.call_args + # Compare the names of the metrics. + assert [m.name for m in mock_kwargs["metrics"]] == [ + vertexai_types.PrebuiltMetric.SAFETY.name + ] diff --git a/tests/unittests/evaluation/test_vertex_ai_eval_facade.py b/tests/unittests/evaluation/test_vertex_ai_eval_facade.py new file mode 100644 index 00000000..63468be0 --- /dev/null +++ b/tests/unittests/evaluation/test_vertex_ai_eval_facade.py @@ -0,0 +1,226 @@ +# 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. + +"""Tests for the Response Evaluator.""" +import random +from unittest.mock import patch + +from google.adk.evaluation.eval_case import Invocation +from google.adk.evaluation.evaluator import EvalStatus +from google.adk.evaluation.vertex_ai_eval_facade import _VertexAiEvalFacade +from google.genai import types as genai_types +import pytest +from vertexai import types as vertexai_types + + +@patch( + "google.adk.evaluation.vertex_ai_eval_facade._VertexAiEvalFacade._perform_eval" +) +class TestVertexAiEvalFacade: + """A class to help organize "patch" that are applicable to all tests.""" + + def test_evaluate_invocations_metric_passed(self, mock_perform_eval): + """Test evaluate_invocations function for a metric.""" + actual_invocations = [ + Invocation( + user_content=genai_types.Content( + parts=[genai_types.Part(text="This is a test query.")] + ), + final_response=genai_types.Content( + parts=[ + genai_types.Part(text="This is a test candidate response.") + ] + ), + ) + ] + expected_invocations = [ + Invocation( + user_content=genai_types.Content( + parts=[genai_types.Part(text="This is a test query.")] + ), + final_response=genai_types.Content( + parts=[genai_types.Part(text="This is a test reference.")] + ), + ) + ] + evaluator = _VertexAiEvalFacade( + threshold=0.8, metric_name=vertexai_types.PrebuiltMetric.COHERENCE + ) + # Mock the return value of _perform_eval + mock_perform_eval.return_value = vertexai_types.EvaluationResult( + summary_metrics=[vertexai_types.AggregatedMetricResult(mean_score=0.9)], + eval_case_results=[], + ) + + evaluation_result = evaluator.evaluate_invocations( + actual_invocations, expected_invocations + ) + + assert evaluation_result.overall_score == 0.9 + assert evaluation_result.overall_eval_status == EvalStatus.PASSED + mock_perform_eval.assert_called_once() + _, mock_kwargs = mock_perform_eval.call_args + # Compare the names of the metrics. + assert [m.name for m in mock_kwargs["metrics"]] == [ + vertexai_types.PrebuiltMetric.COHERENCE.name + ] + + def test_evaluate_invocations_metric_failed(self, mock_perform_eval): + """Test evaluate_invocations function for a metric.""" + actual_invocations = [ + Invocation( + user_content=genai_types.Content( + parts=[genai_types.Part(text="This is a test query.")] + ), + final_response=genai_types.Content( + parts=[ + genai_types.Part(text="This is a test candidate response.") + ] + ), + ) + ] + expected_invocations = [ + Invocation( + user_content=genai_types.Content( + parts=[genai_types.Part(text="This is a test query.")] + ), + final_response=genai_types.Content( + parts=[genai_types.Part(text="This is a test reference.")] + ), + ) + ] + evaluator = _VertexAiEvalFacade( + threshold=0.8, metric_name=vertexai_types.PrebuiltMetric.COHERENCE + ) + # Mock the return value of _perform_eval + mock_perform_eval.return_value = vertexai_types.EvaluationResult( + summary_metrics=[vertexai_types.AggregatedMetricResult(mean_score=0.7)], + eval_case_results=[], + ) + + evaluation_result = evaluator.evaluate_invocations( + actual_invocations, expected_invocations + ) + + assert evaluation_result.overall_score == 0.7 + assert evaluation_result.overall_eval_status == EvalStatus.FAILED + mock_perform_eval.assert_called_once() + _, mock_kwargs = mock_perform_eval.call_args + # Compare the names of the metrics. + assert [m.name for m in mock_kwargs["metrics"]] == [ + vertexai_types.PrebuiltMetric.COHERENCE.name + ] + + def test_evaluate_invocations_metric_no_score(self, mock_perform_eval): + """Test evaluate_invocations function for a metric.""" + actual_invocations = [ + Invocation( + user_content=genai_types.Content( + parts=[genai_types.Part(text="This is a test query.")] + ), + final_response=genai_types.Content( + parts=[ + genai_types.Part(text="This is a test candidate response.") + ] + ), + ) + ] + expected_invocations = [ + Invocation( + user_content=genai_types.Content( + parts=[genai_types.Part(text="This is a test query.")] + ), + final_response=genai_types.Content( + parts=[genai_types.Part(text="This is a test reference.")] + ), + ) + ] + evaluator = _VertexAiEvalFacade( + threshold=0.8, metric_name=vertexai_types.PrebuiltMetric.COHERENCE + ) + # Mock the return value of _perform_eval + mock_perform_eval.return_value = vertexai_types.EvaluationResult( + summary_metrics=[], + eval_case_results=[], + ) + + evaluation_result = evaluator.evaluate_invocations( + actual_invocations, expected_invocations + ) + + assert evaluation_result.overall_score is None + assert evaluation_result.overall_eval_status == EvalStatus.NOT_EVALUATED + mock_perform_eval.assert_called_once() + _, mock_kwargs = mock_perform_eval.call_args + # Compare the names of the metrics. + assert [m.name for m in mock_kwargs["metrics"]] == [ + vertexai_types.PrebuiltMetric.COHERENCE.name + ] + + def test_evaluate_invocations_metric_multiple_invocations( + self, mock_perform_eval + ): + """Test evaluate_invocations function for a metric with multiple invocations.""" + num_invocations = 6 + actual_invocations = [] + expected_invocations = [] + mock_eval_results = [] + random.seed(61553) + scores = [random.random() for _ in range(num_invocations)] + + for i in range(num_invocations): + actual_invocations.append( + Invocation( + user_content=genai_types.Content( + parts=[genai_types.Part(text=f"Query {i+1}")] + ), + final_response=genai_types.Content( + parts=[genai_types.Part(text=f"Response {i+1}")] + ), + ) + ) + expected_invocations.append( + Invocation( + user_content=genai_types.Content( + parts=[genai_types.Part(text=f"Query {i+1}")] + ), + final_response=genai_types.Content( + parts=[genai_types.Part(text=f"Reference {i+1}")] + ), + ) + ) + mock_eval_results.append( + vertexai_types.EvaluationResult( + summary_metrics=[ + vertexai_types.AggregatedMetricResult(mean_score=scores[i]) + ], + eval_case_results=[], + ) + ) + + evaluator = _VertexAiEvalFacade( + threshold=0.8, metric_name=vertexai_types.PrebuiltMetric.COHERENCE + ) + # Mock the return value of _perform_eval + mock_perform_eval.side_effect = mock_eval_results + + evaluation_result = evaluator.evaluate_invocations( + actual_invocations, expected_invocations + ) + + assert evaluation_result.overall_score == pytest.approx( + sum(scores) / num_invocations + ) + assert evaluation_result.overall_eval_status == EvalStatus.FAILED + assert mock_perform_eval.call_count == num_invocations