google-cloud-agentplatform 2.4.0__tar.gz → 2.6.0__tar.gz

{google_cloud_agentplatform-2.4.0/google_cloud_agentplatform.egg-info → google_cloud_agentplatform-2.6.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: google-cloud-agentplatform
-Version: 2.4.0
+Version: 2.6.0
 Summary: Google Cloud Agent Platform Client Library
 Author-email: Google LLC <googleapis-packages@google.com>
 License: Apache-2.0
@@ -31,7 +31,7 @@ Requires-Dist: google-cloud-storage<4.0.0,>=1.32.0; extra == "agent-engines"
 Requires-Dist: opentelemetry-sdk<2; extra == "agent-engines"
 Requires-Dist: opentelemetry-exporter-gcp-logging<2.0.0,>=1.11.0a0; extra == "agent-engines"
 Requires-Dist: opentelemetry-exporter-otlp-proto-http<2; extra == "agent-engines"
-Requires-Dist: opentelemetry-instrumentation-google-genai<1.0.0,>=0.3b0; extra == "agent-engines"
+Requires-Dist: opentelemetry-instrumentation-google-genai<2.0.0,>=0.3b0; extra == "agent-engines"
 Requires-Dist: pydantic<3,>=2.11.1; extra == "agent-engines"
 Provides-Extra: a2a
 Requires-Dist: google-cloud-agentplatform[agent_engines]; extra == "a2a"

{google_cloud_agentplatform-2.4.0 → google_cloud_agentplatform-2.6.0}/agentplatform/_genai/_evals_common.py

@@ -22,6 +22,7 @@ import datetime
 import json
 import logging
 import os
+import random
 import threading
 import time
 from typing import Any, Callable, Literal, Optional, Union, cast
@@ -122,6 +123,11 @@ def _temp_logger_level(logger_name: str, level: int) -> None:  # type: ignore[mi
         logger_instance.setLevel(original_level)
 
 
+def _is_missing(value: Any) -> bool:
+    """Returns True for None or a scalar NaN, such as a missing DataFrame cell."""
+    return bool(pd.api.types.is_scalar(value) and pd.isna(value))
+
+
 def _get_api_client_with_location(
     api_client: BaseApiClient, location: Optional[str]
 ) -> BaseApiClient:
@@ -153,6 +159,8 @@ def _get_runtime_instance(
         client = agentplatform.Client(
             project=api_client.project,
             location=api_client.location,
+            credentials=api_client._credentials,
+            http_options=api_client._http_options,
         )
         _thread_local_data.runtime_instances[agent_name] = client.runtimes.get(
             name=agent_name
@@ -160,6 +168,45 @@ def _get_runtime_instance(
     return _thread_local_data.runtime_instances[agent_name]
 
 
+def _is_retryable_error(e: Exception) -> bool:
+    """Returns True for transient errors that are safe to retry with backoff."""
+    if isinstance(e, genai_errors.APIError):
+        return e.code in _evals_constant.RETRYABLE_STATUS_CODES
+    return isinstance(
+        e, (api_exceptions.ResourceExhausted, api_exceptions.ServiceUnavailable)
+    )
+
+
+def _get_retry_backoff(
+    e: Exception, attempt: int, max_retries: int, operation: str
+) -> Optional[float]:
+    """Returns the seconds to wait before the next attempt, or None to stop.
+
+    Retryable errors back off exponentially with jitter. Other errors fail
+    fast, and the last attempt never waits.
+    """
+    if not _is_retryable_error(e) or attempt == max_retries - 1:
+        logger.error(
+            "Error during %s on attempt %d/%d: %s",
+            operation,
+            attempt + 1,
+            max_retries,
+            e,
+        )
+        return None
+    backoff = 2**attempt + random.uniform(0, 1)
+    logger.warning(
+        "Retryable error during %s on attempt %d/%d: %s. Retrying in %.1f"
+        " seconds...",
+        operation,
+        attempt + 1,
+        max_retries,
+        e,
+        backoff,
+    )
+    return backoff
+
+
 def _generate_content_with_retry(
     api_client: BaseApiClient,
     model: str,
@@ -222,29 +269,11 @@ def _generate_content_with_retry(
                         }
                 else:
                     return response
-        except api_exceptions.ResourceExhausted as e:
-            logger.warning(
-                "Resource Exhausted error on attempt %d/%d: %s. Retrying in %s"
-                " seconds...",
-                attempt + 1,
-                max_retries,
-                e,
-                2**attempt,
-            )
-            if attempt == max_retries - 1:
-                return {"error": f"Resource exhausted after retries: {e}"}
-            time.sleep(2**attempt)
         except Exception as e:  # pylint: disable=broad-exception-caught
-            logger.error(
-                "Unexpected error during generate_content on attempt %d/%d: %s",
-                attempt + 1,
-                max_retries,
-                e,
-            )
-
-            if attempt == max_retries - 1:
-                return {"error": f"Failed after retries: {e}"}
-            time.sleep(1)
+            backoff = _get_retry_backoff(e, attempt, max_retries, "generate_content")
+            if backoff is None:
+                return {"error": f"Failed on attempt {attempt + 1}/{max_retries}: {e}"}
+            time.sleep(backoff)
     return {"error": f"Failed to generate content after {max_retries} retries"}
 
 
@@ -267,9 +296,8 @@ def _build_generate_content_config(
     for key in [
         "system_instruction",
         "tools",
-        "tools_config",
+        "tool_config",
         "safety_settings",
-        "labels",
     ]:
         if key in request_dict:
             merged_config_dict[key] = request_dict[key]
@@ -1561,7 +1589,7 @@ def _build_interaction_id_dataset(
     eval_cases = []
     for i, row in enumerate(loaded_data):
         interaction = row.get(_evals_constant.INTERACTION_ID)
-        if not interaction:
+        if _is_missing(interaction) or not interaction:
             raise ValueError(f"Missing `interaction_id` value for row {i}.")
         eval_cases.append(
             types.EvalCase(
@@ -1871,12 +1899,12 @@ def _execute_inference_concurrently(
     max_workers = AGENT_MAX_WORKERS if runtime or agent else MAX_WORKERS
     with tqdm(total=len(prompt_dataset), desc=progress_desc) as pbar:
         with concurrent.futures.ThreadPoolExecutor(max_workers=max_workers) as executor:
-            for index, row in prompt_dataset.iterrows():
+            for index, (_, row) in enumerate(prompt_dataset.iterrows()):
                 try:
                     if (
                         has_agent_data
                         and AGENT_DATA in row.index
-                        and row.get(AGENT_DATA) is not None
+                        and not _is_missing(row.get(AGENT_DATA))
                     ):
                         agent_data_obj = row[AGENT_DATA]
                         if isinstance(agent_data_obj, dict):
@@ -2101,7 +2129,7 @@ def _run_litellm_inference(
 
     with tqdm(total=len(prompt_dataset), desc=f"LiteLLM Inference ({model})") as pbar:
         with concurrent.futures.ThreadPoolExecutor(max_workers=MAX_WORKERS) as executor:
-            for index, row in prompt_dataset.iterrows():
+            for index, (_, row) in enumerate(prompt_dataset.iterrows()):
                 messages = _convert_prompt_row_to_litellm_messages(row)
                 future = executor.submit(
                     _call_litellm_completion, model=model, messages=messages
@@ -3481,7 +3509,9 @@ def _create_runtime_session(
             raise RuntimeError(
                 "Failed to create session: runtime.api_client is None."
             ) from exc
-        operation = runtime.api_client.sessions.create(
+        from . import sessions
+
+        operation = sessions.Sessions(runtime.api_client._api_client).create(
             name=runtime.api_resource.name,
             user_id=user_id,
             config=types.CreateRuntimeSessionConfig(
@@ -3545,10 +3575,12 @@ def _execute_agent_run_with_retry(
             return {"error": "runtime.api_resource is None."}
         if runtime.api_client is None:
             return {"error": "runtime.api_client is None."}
+        from . import sessions
+
         session_name = f"{runtime.api_resource.name}/sessions/{session_id}"
         base_ts = datetime.datetime(2000, 1, 1, tzinfo=datetime.timezone.utc)
         for i, ag_event in enumerate(history_events):
-            runtime.api_client.sessions.events.append(
+            sessions.Sessions(runtime.api_client._api_client).events.append(
                 name=session_name,
                 author=ag_event.author or "user",
                 invocation_id="history",
@@ -3570,28 +3602,11 @@ def _execute_agent_run_with_retry(
                 if event and CONTENT in event and PARTS in event[CONTENT]:
                     responses.append(event)
             return responses
-        except api_exceptions.ResourceExhausted as e:
-            logger.warning(
-                "Resource Exhausted error on attempt %d/%d: %s. Retrying in %s"
-                " seconds...",
-                attempt + 1,
-                max_retries,
-                e,
-                2**attempt,
-            )
-            if attempt == max_retries - 1:
-                return {"error": f"Resource exhausted after retries: {e}"}
-            time.sleep(2**attempt)
         except Exception as e:  # pylint: disable=broad-exception-caught
-            logger.error(
-                "Unexpected error during agent engine run on attempt %d/%d: %s",
-                attempt + 1,
-                max_retries,
-                e,
-            )
-            if attempt == max_retries - 1:
-                return {"error": f"Failed after retries: {e}"}
-            time.sleep(1)
+            backoff = _get_retry_backoff(e, attempt, max_retries, "agent engine run")
+            if backoff is None:
+                return {"error": f"Failed on attempt {attempt + 1}/{max_retries}: {e}"}
+            time.sleep(backoff)
     return {"error": f"Failed to get agent run results after {max_retries} retries"}
 
 
@@ -3623,7 +3638,7 @@ async def _execute_local_agent_run_with_retry_async(
     # Lazy-import ADK dependencies to avoid top-level import failures when
     # google-adk is not installed.
     from google.adk.runners import Runner
-    from google.adk.sessions import InMemorySessionService
+    from google.adk.sessions.in_memory_session_service import InMemorySessionService
 
     # Multi-turn agent scraping with user simulation.
     if user_simulator_config or "conversation_plan" in row:
@@ -3697,28 +3712,13 @@ async def _execute_local_agent_run_with_retry_async(
                     if event and CONTENT in event and PARTS in event[CONTENT]:
                         events.append(event)
                 return events
-            except api_exceptions.ResourceExhausted as e:
-                logger.warning(
-                    "Resource Exhausted error on attempt %d/%d: %s. Retrying"
-                    " in %s seconds...",
-                    attempt + 1,
-                    max_retries,
-                    e,
-                    2**attempt,
-                )
-                if attempt == max_retries - 1:
-                    return {"error": f"Resource exhausted after retries: {e}"}
-                await asyncio.sleep(2**attempt)
             except Exception as e:  # pylint: disable=broad-exception-caught
-                logger.error(
-                    "Unexpected error during agent run on attempt %d/%d: %s",
-                    attempt + 1,
-                    max_retries,
-                    e,
-                )
-                if attempt == max_retries - 1:
-                    return {"error": f"Failed after retries: {e}"}
-                await asyncio.sleep(1)
+                backoff = _get_retry_backoff(e, attempt, max_retries, "agent run")
+                if backoff is None:
+                    return {
+                        "error": f"Failed on attempt {attempt + 1}/{max_retries}: {e}"
+                    }
+                await asyncio.sleep(backoff)
         return {"error": f"Failed to get agent run results after {max_retries} retries"}
 
 
@@ -3734,7 +3734,10 @@ def _convert_gcs_to_evaluation_item_result(
         return types.EvaluationItemResult(**eval_item_data)
     except Exception as e:
         logger.error(
-            "Failed to load evaluation result from GCS: %s. Error: %s", gcs_uri, e
+            "Failed to load evaluation result from GCS: %s. Error: %s",
+            gcs_uri,
+            e,
+            exc_info=True,
         )
     return types.EvaluationItemResult()
 
@@ -3751,7 +3754,10 @@ def _convert_gcs_to_evaluation_item_request(
         return types.EvaluationItemRequest(**eval_item_data)
     except Exception as e:
         logger.error(
-            "Failed to load evaluation request from GCS: %s. Error: %s", gcs_uri, e
+            "Failed to load evaluation request from GCS: %s. Error: %s",
+            gcs_uri,
+            e,
+            exc_info=True,
         )
     return types.EvaluationItemRequest()
 
@@ -3874,7 +3880,7 @@ def _drop_empty_columns(df: "pd.DataFrame") -> "pd.DataFrame":
     def is_empty(x: Any) -> bool:
         if isinstance(x, (list, dict)):
             return not x
-        return pd.isna(x)  # type: ignore[no-any-return]
+        return _is_missing(x)
 
     cols_to_drop = [col for col in df.columns if df[col].apply(is_empty).all()]
     return df.drop(columns=cols_to_drop)
@@ -3925,16 +3931,16 @@ def _get_eval_cases_eval_dfs_from_eval_items(
           item.
         - eval_dfs: A list of EvaluationDatasets, one for each candidate.
     """
-    dataset_rows = []
+    dataset_rows: list[dict[str, Any]] = []
     eval_case_results = []
-    for index, eval_item in enumerate(eval_items):
+    for eval_item in eval_items:
         if (
             eval_item
             and eval_item.evaluation_response
             and eval_item.evaluation_response.request
         ):
             eval_case_results.append(
-                _get_eval_case_result_from_eval_item(index, eval_item)
+                _get_eval_case_result_from_eval_item(len(dataset_rows), eval_item)
             )
             dataset_rows.append(
                 _convert_request_to_dataset_row(eval_item.evaluation_response.request)
@@ -4119,6 +4125,36 @@ def _get_content(row: dict[str, Any], column: str) -> Optional[genai_types.Conte
         )
 
 
+# Columns mapped to EvaluationItemRequest fields; all others go to extra_context.
+_EVALUATION_SET_STANDARD_COLUMNS = _evals_constant.COMMON_DATASET_COLUMNS | {
+    _evals_constant.RESPONSE,
+    "interaction",
+    "gemini_agent",
+}
+
+
+def _to_extra_context_content(value: Any) -> Optional[genai_types.Content]:
+    if isinstance(value, str):
+        return genai_types.Content(parts=[genai_types.Part(text=value)])
+    if isinstance(value, genai_types.Content):
+        return value
+    parts = []
+    for content in _evals_metric_handlers._value_to_content_list(value):
+        parts.extend(content.parts or [])
+    return genai_types.Content(parts=parts) if parts else None
+
+
+def _get_extra_context(row: "pd.Series") -> Optional[dict[str, genai_types.Content]]:
+    extra_context = {}
+    for column, value in row.items():
+        if column in _EVALUATION_SET_STANDARD_COLUMNS or _is_missing(value):
+            continue
+        content = _to_extra_context_content(value)
+        if content:
+            extra_context[str(column)] = content
+    return extra_context or None
+
+
 def _create_evaluation_set_from_dataframe(
     api_client: BaseApiClient,
     gcs_dest_prefix: str,
@@ -4173,6 +4209,8 @@ def _create_evaluation_set_from_dataframe(
         if _evals_constant.RESPONSE in row or agent_data_obj or intermediate_events:
             # Resolve the oneof conflict: prioritize agent_data over flat text
             response_text = row.get(_evals_constant.RESPONSE) or None
+            if _is_missing(response_text):
+                response_text = None
 
             if agent_data_obj and response_text:
                 logger.info(
@@ -4227,7 +4265,7 @@ def _create_evaluation_set_from_dataframe(
             prompt = types.EvaluationPrompt(
                 prompt_template_data=types.PromptTemplateData(values=values)
             )
-        elif _evals_constant.PROMPT in row:
+        elif not _is_missing(row.get(_evals_constant.PROMPT)):
             prompt = types.EvaluationPrompt(text=row[_evals_constant.PROMPT])
 
         eval_item_requests.append(
@@ -4235,12 +4273,13 @@ def _create_evaluation_set_from_dataframe(
                 prompt=prompt or None,
                 golden_response=(
                     types.CandidateResponse(text=row[_evals_constant.REFERENCE])
-                    if _evals_constant.REFERENCE in row
+                    if not _is_missing(row.get(_evals_constant.REFERENCE))
                     else None
                 ),
                 candidate_responses=(
                     candidate_responses if candidate_responses else None
                 ),
+                extra_context=_get_extra_context(row),
             )
         )
     logger.info("Writing evaluation item requests to GCS.")

{google_cloud_agentplatform-2.4.0 → google_cloud_agentplatform-2.6.0}/agentplatform/_genai/_evals_constant.py

@@ -120,3 +120,17 @@ COMMON_DATASET_COLUMNS = frozenset(
         INTERACTION_ID,
     }
 )
+
+# HTTP status codes that are safe to retry with backoff.
+RETRYABLE_STATUS_CODES = frozenset(
+    {
+        408,  # RequestTimeout (DEADLINE_EXCEEDED)
+        409,  # Conflict / Aborted (ABORTED)
+        429,  # TooManyRequests / ResourceExhausted (RESOURCE_EXHAUSTED)
+        499,  # Client Closed Request (CANCELLED)
+        500,  # InternalServerError (INTERNAL)
+        502,  # BadGateway
+        503,  # ServiceUnavailable (UNAVAILABLE)
+        504,  # GatewayTimeout (DEADLINE_EXCEEDED)
+    }
+)

{google_cloud_agentplatform-2.4.0 → google_cloud_agentplatform-2.6.0}/agentplatform/_genai/_evals_data_converters.py

@@ -230,6 +230,18 @@ class _FlattenEvalDataConverter(_evals_utils.EvalDataConverter):
                     "Invalid prompt type for case %s: %s" % (i, type(prompt_data))
                 )
 
+            if (
+                isinstance(conversation_history_data, str)
+                and conversation_history_data.strip()
+            ):
+                try:
+                    conversation_history_data = json.loads(conversation_history_data)
+                except json.JSONDecodeError:
+                    logger.warning(
+                        "Could not decode conversation_history JSON string for case %s.",
+                        eval_case_id,
+                    )
+
             conversation_history: Optional[list[types.evals.Message]] = None
             if isinstance(conversation_history_data, list):
                 conversation_history = []
@@ -242,6 +254,11 @@ class _FlattenEvalDataConverter(_evals_utils.EvalDataConverter):
                             )
                         )
                     elif isinstance(content, dict):
+                        if isinstance(content.get("content"), str):
+                            content = {
+                                "role": content.get("role"),
+                                "parts": [{"text": content["content"]}],
+                            }
                         try:
                             validated_content = genai_types.Content.model_validate(
                                 content

{google_cloud_agentplatform-2.4.0 → google_cloud_agentplatform-2.6.0}/agentplatform/_genai/_evals_metric_handlers.py

@@ -39,19 +39,6 @@ from . import types
 
 logger = logging.getLogger(__name__)
 _MAX_RETRIES = 5
-# HTTP status codes that are safe to retry with backoff.
-_RETRYABLE_STATUS_CODES = frozenset(
-    {
-        408,  # RequestTimeout (DEADLINE_EXCEEDED)
-        409,  # Conflict / Aborted (ABORTED)
-        429,  # TooManyRequests / ResourceExhausted (RESOURCE_EXHAUSTED)
-        499,  # Client Closed Request (CANCELLED)
-        500,  # InternalServerError (INTERNAL)
-        502,  # BadGateway
-        503,  # ServiceUnavailable (UNAVAILABLE)
-        504,  # GatewayTimeout (DEADLINE_EXCEEDED)
-    }
-)
 
 R = TypeVar("R")
 T = TypeVar("T", types.Metric, types.MetricSource, types.LLMMetric)
@@ -64,7 +51,7 @@ def _call_with_retry(
     """Calls ``fn()`` with exponential backoff + jitter on retryable errors.
 
     Retries up to ``_MAX_RETRIES`` times on errors whose HTTP status code is
-    in ``_RETRYABLE_STATUS_CODES`` (Aborted, DeadlineExceeded,
+    in ``_evals_constant.RETRYABLE_STATUS_CODES`` (Aborted, DeadlineExceeded,
     ResourceExhausted, ServiceUnavailable, Cancelled). Non-retryable errors
     are re-raised immediately. If all retries are exhausted the last
     exception is re-raised so the caller can decide how to handle it.
@@ -84,7 +71,7 @@ def _call_with_retry(
         try:
             return fn()
         except genai_errors.APIError as e:
-            if e.code in _RETRYABLE_STATUS_CODES:
+            if e.code in _evals_constant.RETRYABLE_STATUS_CODES:
                 backoff = 2**attempt + random.uniform(0, 1)
                 logger.warning(
                     "Retryable error (code=%s) on attempt %d/%d for metric"
@@ -686,6 +673,12 @@ def _build_evaluation_instance(
             other_data_map["context"] = types.evals.InstanceData(text=eval_case.context)
         elif isinstance(eval_case.context, genai_types.Content):
             other_data_map["context"] = _content_to_instance_data(eval_case.context)
+        else:
+            other_data_map["context"] = types.evals.InstanceData(
+                contents=types.evals.InstanceDataContents(
+                    contents=_value_to_content_list(eval_case.context)
+                )
+            )
 
     # 3. Extract custom variables from LLMMetric templates
     if prompt_template:

{google_cloud_agentplatform-2.4.0 → google_cloud_agentplatform-2.6.0}/agentplatform/_genai/_evals_visualization.py

@@ -23,7 +23,7 @@ import textwrap
 from typing import Any, Optional
 
 import pandas as pd
-from pydantic import errors
+import pydantic_core
 
 from . import _evals_common
 from . import types
@@ -229,6 +229,20 @@ def _extract_dataset_rows(dataset: types.EvaluationDataset) -> list[dict[str, An
             }
             processed_rows.append(processed_row)
 
+    for row in processed_rows:
+        if isinstance(row["agent_data"], str) or not isinstance(
+            row["intermediate_events"], str
+        ):
+            continue
+        try:
+            events = json.loads(row["intermediate_events"])
+        except json.JSONDecodeError:
+            continue
+        if isinstance(events, list) and events:
+            row["agent_data"] = json.dumps(
+                {"turns": [{"turn_index": 0, "events": events}]}, ensure_ascii=False
+            )
+
     return processed_rows
 
 
@@ -1155,7 +1169,7 @@ def get_comparison_html(eval_result_json: str) -> str:
                     winRateText = m.win_rates.map((rate, i) => `<b>${{candidateNames ? candidateNames[i] : `Candidate #${{i+1}}`}}</b> wins: <b>${{(rate * 100).toFixed(1)}}%</b>`).join('<br>');
                     if (m.tie_rate !== undefined) {{ winRateText += `<br>Ties: <b>${{(m.tie_rate * 100).toFixed(1)}}%</b>`; }}
                 }}
-                table += `<tr><td>${{m.metric_name}}</td><td>${{m.mean_score.toFixed(4)}}</td><td>${{m.stdev_score.toFixed(4)}}</td><td>${{winRateText}}</td></tr>`;
+                table += `<tr><td>${{m.metric_name}}</td><td>${{m.mean_score != null ? m.mean_score.toFixed(4) : 'N/A'}}</td><td>${{m.stdev_score != null ? m.stdev_score.toFixed(4) : 'N/A'}}</td><td>${{winRateText}}</td></tr>`;
             }});
             container.innerHTML = table + '</tbody></table>';
         }}
@@ -1469,7 +1483,7 @@ def display_evaluation_result(
         result_dump = eval_result_obj.model_dump(
             mode="json", exclude_none=True, exclude={"evaluation_dataset"}
         )
-    except errors.PydanticSerializationError as e:
+    except pydantic_core.PydanticSerializationError as e:
         logger.error(
             "Serialization Error: %s\nCould not display the evaluation "
             "result due to a data serialization issue. Please check the "

{google_cloud_agentplatform-2.4.0 → google_cloud_agentplatform-2.6.0}/agentplatform/_genai/_runtimes_utils.py

@@ -52,6 +52,7 @@ import httpx
 import proto
 
 from google.api_core import exceptions
+from google.genai import errors as google_genai_errors
 from google.genai import types as google_genai_types
 from google.api import httpbody_pb2
 from google.protobuf import struct_pb2
@@ -564,6 +565,23 @@ def _await_operation(
     return operation
 
 
+def _raise_for_operation_error(
+    *,
+    operation: Any,
+    message: str,
+) -> None:
+    """Raises an APIError for a failed operation with a context note."""
+    try:
+        google_genai_errors.APIError.raise_error(
+            operation.error.get("code"), operation.error, None
+        )
+    except google_genai_errors.APIError as e:
+        # TODO(yifany): Drop guard once Python 3.10 support is removed.
+        if sys.version_info >= (3, 11):
+            e.add_note(message)
+        raise
+
+
 def _compare_requirements(
     *,
     requirements: Mapping[str, str],

{google_cloud_agentplatform-2.4.0 → google_cloud_agentplatform-2.6.0}/agentplatform/_genai/evals.py

@@ -3070,7 +3070,7 @@ class Evals(_api_module.BaseModule):
                 events are replayed as session history for local ADK agents.
           model: Optional type is experimental and may change in future versions.
                 The model to use for inference, optional for agent evaluations.
-              - For Google Gemini models, provide the model name string (e.g., "gemini-2.5-flash").
+              - For Google Gemini models, provide the model name string (e.g., "gemini-flash-latest").
               - For third-party models via LiteLLM, use the format "provider/model_name"
                 (e.g., "openai/gpt-4o"). Ensure the necessary API key (e.g., OPENAI_API_KEY)
                 is set as an environment variable.
@@ -3342,7 +3342,7 @@ class Evals(_api_module.BaseModule):
               with the content.
             generator_model_config: Optional. Configuration for the model used
               in custom rubric generation. Only used if `prompt_template` is provided.
-              e.g., {"autorater_model": "gemini-2.5-flash"}.
+              e.g., {"autorater_model": "gemini-flash-latest"}.
             rubric_content_type: Optional. The type of rubric content to be
               generated. Only used if `prompt_template` is provided.
             rubric_type_ontology: Optional. A pre-defined list of allowed types
@@ -3619,12 +3619,12 @@ class Evals(_api_module.BaseModule):
                If provided, `agent_info` must be None. If omitted and `agent_info` is provided,
                this will be automatically constructed using `agent_info` and `user_simulator_config`.
                The `model` field of an inference config accepts a short Gemini model
-               name (e.g. `gemini-2.5-flash`), which is automatically expanded to a
+               name (e.g. `gemini-flash-latest`), which is automatically expanded to a
                fully-qualified resource name using the client's project and location,
                or an already fully-qualified publisher-model or endpoint resource
                name.
                Example:
-               {"candidate-1": types.EvaluationRunInferenceConfig(model="gemini-2.5-flash")}
+               {"candidate-1": types.EvaluationRunInferenceConfig(model="gemini-flash-latest")}
            labels: The labels to apply to the evaluation run.
            loss_analysis_metrics: This field is experimental and may change in future
                versions. Optional list of metrics to run loss analysis on. The
@@ -6261,12 +6261,12 @@ class AsyncEvals(_api_module.BaseModule):
               If provided, `agent_info` must be None. If omitted and `agent_info` is provided,
               this will be automatically constructed using `agent_info` and `user_simulator_config`.
               The `model` field of an inference config accepts a short Gemini model
-              name (e.g. `gemini-2.5-flash`), which is automatically expanded to a
+              name (e.g. `gemini-flash-latest`), which is automatically expanded to a
               fully-qualified resource name using the client's project and location,
               or an already fully-qualified publisher-model or endpoint resource
               name.
               Example:
-              {"candidate-1": types.EvaluationRunInferenceConfig(model="gemini-2.5-flash")}
+              {"candidate-1": types.EvaluationRunInferenceConfig(model="gemini-flash-latest")}
           red_teaming_config: This field is experimental and may change in future
               versions. Optional configuration for automated Agent Red Teaming
               analysis. Specifies attack categories and vulnerable tools to