encounter/adk-python
0
0
Fork 0
mirror of https://github.com/encounter/adk-python.git synced 2026-07-09 18:19:28 -07:00
Code Issues Packages Projects Releases Wiki Activity

feat: Add Safety evaluator metric

Browse Source 
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
This commit is contained in:
Ankur Sharma
2025-07-02 11:30:31 -07:00
committed by Copybara-Service
parent 62c4a85917
commit 0bd05df471
10 changed files with 544 additions and 230 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/unittests/evaluation/test_response_evaluator.py
+5 -144
View File
@@ -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
tests/unittests/evaluation/test_safety_evaluator.py
+78
View File
@@ -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
]
tests/unittests/evaluation/test_vertex_ai_eval_facade.py
+226
View File
@@ -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