chore: Add default retry options as fall back to llm_request that are made during evals

In order to make evals more resilient to temporary model failures, we add retry options to llm_requests that are made during evals.

Note that this is a fall back option, if the developer has already specified their own retry options, then those will be honored.

Co-authored-by: Ankur Sharma <ankusharma@google.com>
PiperOrigin-RevId: 832383755

src/google/adk/evaluation/_retry_options_utils.py

@@ -0,0 +1,75 @@
+# 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 import Optional
+
+from google.genai import types
+from typing_extensions import override
+
+from ..agents.callback_context import CallbackContext
+from ..models.llm_request import LlmRequest
+from ..models.llm_response import LlmResponse
+from ..plugins.base_plugin import BasePlugin
+
+_RETRY_HTTP_STATUS_CODES = (
+    408,  # Request timeout.
+    429,  # Too many requests.
+    500,  # Internal server error.
+    502,  # Bad gateway.
+    503,  # Service unavailable.
+    504,  # Gateway timeout
+)
+_DEFAULT_HTTP_RETRY_OPTIONS = types.HttpRetryOptions(
+    attempts=7,
+    initial_delay=5.0,
+    max_delay=120,
+    exp_base=2.0,
+    http_status_codes=_RETRY_HTTP_STATUS_CODES,
+)
+
+
+def add_default_retry_options_if_not_present(llm_request: LlmRequest):
+  """Adds default HTTP Retry Options, if they are not present on the llm_request.
+
+  NOTE: This implementation is intended for eval systems internal usage. Do not
+  take direct dependency on it.
+  """
+  llm_request.config = llm_request.config or types.GenerateContentConfig()
+
+  llm_request.config.http_options = (
+      llm_request.config.http_options or types.HttpOptions()
+  )
+  llm_request.config.http_options.retry_options = (
+      llm_request.config.http_options.retry_options
+      or _DEFAULT_HTTP_RETRY_OPTIONS
+  )
+
+
+class EnsureRetryOptionsPlugin(BasePlugin):
+  """This plugin adds retry options to llm_request, if they are not present.
+
+  This is done to ensure that temporary outages with the model provider don't
+  affect eval runs.
+
+  NOTE: This implementation is intended for eval systems internal usage. Do not
+  take direct dependency on it.
+  """
+
+  @override
+  async def before_model_callback(
+      self, *, callback_context: CallbackContext, llm_request: LlmRequest
+  ) -> Optional[LlmResponse]:
+    add_default_retry_options_if_not_present(llm_request)

src/google/adk/evaluation/evaluation_generator.py

@@ -35,6 +35,7 @@ from ..sessions.base_session_service import BaseSessionService
 from ..sessions.in_memory_session_service import InMemorySessionService
 from ..sessions.session import Session
 from ..utils.context_utils import Aclosing
+from ._retry_options_utils import EnsureRetryOptionsPlugin
 from .app_details import AgentDetails
 from .app_details import AppDetails
 from .eval_case import EvalCase
@@ -225,13 +226,19 @@ class EvaluationGenerator:
     request_intercepter_plugin = _RequestIntercepterPlugin(
         name="request_intercepter_plugin"
     )
+    # We ensure that there is some kind of retries on the llm_requests that are
+    # generated from the Agent. This is done to make inferencing step of evals
+    # more resilient to temporary model failures.
+    ensure_retry_options_plugin = EnsureRetryOptionsPlugin(
+        name="ensure_retry_options"
+    )
     async with Runner(
         app_name=app_name,
         agent=root_agent,
         artifact_service=artifact_service,
         session_service=session_service,
         memory_service=memory_service,
-        plugins=[request_intercepter_plugin],
+        plugins=[request_intercepter_plugin, ensure_retry_options_plugin],
     ) as runner:
       events = []
       while True:

src/google/adk/evaluation/hallucinations_v1.py

@@ -32,6 +32,7 @@ from ..models.llm_response import LlmResponse
 from ..models.registry import LLMRegistry
 from ..utils.context_utils import Aclosing
 from ..utils.feature_decorator import experimental
+from ._retry_options_utils import add_default_retry_options_if_not_present
 from .app_details import AppDetails
 from .eval_case import Invocation
 from .eval_case import InvocationEvent
@@ -526,6 +527,7 @@ class HallucinationsV1Evaluator(Evaluator):
         ],
         config=self._model_config,
     )
+    add_default_retry_options_if_not_present(segmenter_llm_request)
     try:
       async with Aclosing(
           self._judge_model.generate_content_async(segmenter_llm_request)
@@ -559,6 +561,7 @@ class HallucinationsV1Evaluator(Evaluator):
         ],
         config=self._model_config,
     )
+    add_default_retry_options_if_not_present(validator_llm_request)
     try:
       async with Aclosing(
           self._judge_model.generate_content_async(validator_llm_request)

src/google/adk/evaluation/llm_as_judge.py

@@ -27,6 +27,7 @@ from ..models.llm_response import LlmResponse
 from ..models.registry import LLMRegistry
 from ..utils.context_utils import Aclosing
 from ..utils.feature_decorator import experimental
+from ._retry_options_utils import add_default_retry_options_if_not_present
 from .common import EvalBaseModel
 from .eval_case import Invocation
 from .eval_metrics import BaseCriterion
@@ -142,6 +143,7 @@ class LlmAsJudge(Evaluator):
           ],
           config=self._judge_model_options.judge_model_config,
       )
+      add_default_retry_options_if_not_present(llm_request)
       num_samples = self._judge_model_options.num_samples
       invocation_result_samples = []
       for _ in range(num_samples):

src/google/adk/evaluation/llm_backed_user_simulator.py

@@ -27,6 +27,7 @@ from ..models.llm_request import LlmRequest
 from ..models.registry import LLMRegistry
 from ..utils.context_utils import Aclosing
 from ..utils.feature_decorator import experimental
+from ._retry_options_utils import add_default_retry_options_if_not_present
 from .conversation_scenarios import ConversationScenario
 from .evaluator import Evaluator
 from .user_simulator import BaseUserSimulatorConfig
@@ -200,6 +201,7 @@ class LlmBackedUserSimulator(UserSimulator):
             ),
         ],
     )
+    add_default_retry_options_if_not_present(llm_request)
 
     response = ""
     async with Aclosing(self._llm.generate_content_async(llm_request)) as agen: