azure-ai-evaluation 1.7.0__py3-none-any.whl → 1.9.0__py3-none-any.whl

azure/ai/evaluation/_common/onedp/operations/_patch.py

@@ -1,21 +1,21 @@
-# coding=utf-8
-# --------------------------------------------------------------------------
-# Copyright (c) Microsoft Corporation. All rights reserved.
-# Licensed under the MIT License. See License.txt in the project root for license information.
-# --------------------------------------------------------------------------
-"""Customize generated code here.
-
-Follow our quickstart for examples: https://aka.ms/azsdk/python/dpcodegen/python/customize
-"""
-from typing import List
-
-__all__: List[str] = []  # Add all objects you want publicly available to users at this package level
-
-
-def patch_sdk():
-    """Do not remove from this file.
-
-    `patch_sdk` is a last resort escape hatch that allows you to do customizations
-    you can't accomplish using the techniques described in
-    https://aka.ms/azsdk/python/dpcodegen/python/customize
-    """
+# coding=utf-8
+# --------------------------------------------------------------------------
+# Copyright (c) Microsoft Corporation. All rights reserved.
+# Licensed under the MIT License. See License.txt in the project root for license information.
+# --------------------------------------------------------------------------
+"""Customize generated code here.
+
+Follow our quickstart for examples: https://aka.ms/azsdk/python/dpcodegen/python/customize
+"""
+from typing import List
+
+__all__: List[str] = []  # Add all objects you want publicly available to users at this package level
+
+
+def patch_sdk():
+    """Do not remove from this file.
+
+    `patch_sdk` is a last resort escape hatch that allows you to do customizations
+    you can't accomplish using the techniques described in
+    https://aka.ms/azsdk/python/dpcodegen/python/customize
+    """

azure/ai/evaluation/_common/rai_service.py

@@ -21,10 +21,11 @@ from azure.ai.evaluation._legacy._adapters._errors import MissingRequiredPackage
 from azure.ai.evaluation._exceptions import ErrorBlame, ErrorCategory, ErrorTarget, EvaluationException
 from azure.ai.evaluation._http_utils import AsyncHttpPipeline, get_async_http_client
 from azure.ai.evaluation._model_configurations import AzureAIProject
+from azure.ai.evaluation._user_agent import UserAgentSingleton
 from azure.ai.evaluation._common.utils import is_onedp_project
 from azure.core.credentials import TokenCredential
 from azure.core.exceptions import HttpResponseError
-from azure.core.pipeline.policies import AsyncRetryPolicy
+from azure.core.pipeline.policies import AsyncRetryPolicy, UserAgentPolicy
 
 from .constants import (
     CommonConstants,
@@ -35,20 +36,16 @@ from .constants import (
 )
 from .utils import get_harm_severity_level, retrieve_content_type
 
-try:
-    version = importlib.metadata.version("azure-ai-evaluation")
-except importlib.metadata.PackageNotFoundError:
-    version = "unknown"
-USER_AGENT = "{}/{}".format("azure-ai-evaluation", version)
 
 USER_TEXT_TEMPLATE_DICT: Dict[str, Template] = {
     "DEFAULT": Template("<Human>{$query}</><System>{$response}</>"),
 }
-ML_WORKSPACE  = "https://management.azure.com/.default"
+ML_WORKSPACE = "https://management.azure.com/.default"
 COG_SRV_WORKSPACE = "https://ai.azure.com/.default"
 
 INFERENCE_OF_SENSITIVE_ATTRIBUTES = "inference_sensitive_attributes"
 
+
 def get_formatted_template(data: dict, annotation_task: str) -> str:
     """Given the task and input data, produce a formatted string that will serve as the main
     payload for the RAI service. Requires specific per-task logic.
@@ -71,16 +68,13 @@ def get_formatted_template(data: dict, annotation_task: str) -> str:
         }
         return json.dumps(as_dict)
     if annotation_task == Tasks.CODE_VULNERABILITY:
-        as_dict = {
-            "context": data.get("query", ""),
-            "completion": data.get("response", "")
-        }
+        as_dict = {"context": data.get("query", ""), "completion": data.get("response", "")}
         return json.dumps(as_dict)
     if annotation_task == Tasks.UNGROUNDED_ATTRIBUTES:
         as_dict = {
             "query": data.get("query", ""),
             "response": data.get("response", ""),
-            "context": data.get("context", "")
+            "context": data.get("context", ""),
         }
         return json.dumps(as_dict)
     as_dict = {
@@ -101,7 +95,11 @@ def get_common_headers(token: str, evaluator_name: Optional[str] = None) -> Dict
     :return: The common headers.
     :rtype: Dict
     """
-    user_agent = f"{USER_AGENT} (type=evaluator; subtype={evaluator_name})" if evaluator_name else USER_AGENT
+    user_agent = (
+        f"{UserAgentSingleton().value} (type=evaluator; subtype={evaluator_name})"
+        if evaluator_name
+        else UserAgentSingleton().value
+    )
     return {
         "Authorization": f"Bearer {token}",
         "User-Agent": user_agent,
@@ -113,7 +111,10 @@ def get_async_http_client_with_timeout() -> AsyncHttpPipeline:
         retry_policy=AsyncRetryPolicy(timeout=CommonConstants.DEFAULT_HTTP_TIMEOUT)
     )
 
-async def ensure_service_availability_onedp(client: AIProjectClient, token: str, capability: Optional[str] = None) -> None:
+
+async def ensure_service_availability_onedp(
+    client: AIProjectClient, token: str, capability: Optional[str] = None
+) -> None:
     """Check if the Responsible AI service is available in the region and has the required capability, if relevant.
 
     :param client: The AI project client.
@@ -126,7 +127,7 @@ async def ensure_service_availability_onedp(client: AIProjectClient, token: str,
     """
     headers = get_common_headers(token)
     capabilities = client.evaluations.check_annotation(headers=headers)
-    
+
     if capability and capability not in capabilities:
         msg = f"The needed capability '{capability}' is not supported by the RAI service in this region."
         raise EvaluationException(
@@ -137,7 +138,8 @@ async def ensure_service_availability_onedp(client: AIProjectClient, token: str,
             blame=ErrorBlame.USER_ERROR,
             tsg_link="https://aka.ms/azsdk/python/evaluation/safetyevaluator/troubleshoot",
         )
-    
+
+
 async def ensure_service_availability(rai_svc_url: str, token: str, capability: Optional[str] = None) -> None:
     """Check if the Responsible AI service is available in the region and has the required capability, if relevant.
 
@@ -257,12 +259,13 @@ async def submit_request(
 
 
 async def submit_request_onedp(
-    client: AIProjectClient, 
-    data: dict, 
-    metric: str, 
-    token: str, 
-    annotation_task: str, 
-    evaluator_name: str
+    client: AIProjectClient,
+    data: dict,
+    metric: str,
+    token: str,
+    annotation_task: str,
+    evaluator_name: str,
+    scan_session_id: Optional[str] = None,
 ) -> str:
     """Submit request to Responsible AI service for evaluation and return operation ID
 
@@ -278,12 +281,16 @@ async def submit_request_onedp(
     :type annotation_task: str
     :param evaluator_name: The evaluator name.
     :type evaluator_name: str
+    :param scan_session_id: The scan session ID to use for the evaluation.
+    :type scan_session_id: Optional[str]
     :return: The operation ID.
     :rtype: str
     """
     normalized_user_text = get_formatted_template(data, annotation_task)
     payload = generate_payload(normalized_user_text, metric, annotation_task=annotation_task)
     headers = get_common_headers(token, evaluator_name)
+    if scan_session_id:
+        headers["client_request_id"] = scan_session_id
     response = client.evaluations.submit_annotation(payload, headers=headers)
     result = json.loads(response)
     operation_id = result["location"].split("/")[-1]
@@ -326,6 +333,7 @@ async def fetch_result(operation_id: str, rai_svc_url: str, credential: TokenCre
         sleep_time = RAIService.SLEEP_TIME**request_count
         await asyncio.sleep(sleep_time)
 
+
 async def fetch_result_onedp(client: AIProjectClient, operation_id: str, token: str) -> Dict:
     """Fetch the annotation result from Responsible AI service
 
@@ -349,11 +357,14 @@ async def fetch_result_onedp(client: AIProjectClient, operation_id: str, token:
             request_count += 1
             time_elapsed = time.time() - start
             if time_elapsed > RAIService.TIMEOUT:
-                raise TimeoutError(f"Fetching annotation result {request_count} times out after {time_elapsed:.2f} seconds")
+                raise TimeoutError(
+                    f"Fetching annotation result {request_count} times out after {time_elapsed:.2f} seconds"
+                )
 
             sleep_time = RAIService.SLEEP_TIME**request_count
             await asyncio.sleep(sleep_time)
 
+
 def parse_response(  # pylint: disable=too-many-branches,too-many-statements
     batch_response: List[Dict], metric_name: str, metric_display_name: Optional[str] = None
 ) -> Dict[str, Union[str, float]]:
@@ -382,10 +393,13 @@ def parse_response(  # pylint: disable=too-many-branches,too-many-statements
         result = {}
         if not batch_response or len(batch_response[0]) == 0:
             return {}
-        if metric_name == EvaluationMetrics.UNGROUNDED_ATTRIBUTES and INFERENCE_OF_SENSITIVE_ATTRIBUTES in batch_response[0]:
-            batch_response[0] = { 
-                EvaluationMetrics.UNGROUNDED_ATTRIBUTES: batch_response[0][INFERENCE_OF_SENSITIVE_ATTRIBUTES] 
-            } 
+        if (
+            metric_name == EvaluationMetrics.UNGROUNDED_ATTRIBUTES
+            and INFERENCE_OF_SENSITIVE_ATTRIBUTES in batch_response[0]
+        ):
+            batch_response[0] = {
+                EvaluationMetrics.UNGROUNDED_ATTRIBUTES: batch_response[0][INFERENCE_OF_SENSITIVE_ATTRIBUTES]
+            }
         if metric_name == EvaluationMetrics.PROTECTED_MATERIAL and metric_name not in batch_response[0]:
             pm_metric_names = {"artwork", "fictional_characters", "logos_and_brands"}
             for pm_metric_name in pm_metric_names:
@@ -421,7 +435,10 @@ def parse_response(  # pylint: disable=too-many-branches,too-many-statements
             result[metric_display_name + "_information_gathering"] = (
                 parsed_response["information_gathering"] if "information_gathering" in parsed_response else math.nan
             )
-        if metric_name == EvaluationMetrics.CODE_VULNERABILITY or metric_name == EvaluationMetrics.UNGROUNDED_ATTRIBUTES:
+        if (
+            metric_name == EvaluationMetrics.CODE_VULNERABILITY
+            or metric_name == EvaluationMetrics.UNGROUNDED_ATTRIBUTES
+        ):
             # Add all attributes under the details.
             details = {}
             for key, value in parsed_response.items():
@@ -587,7 +604,9 @@ async def get_rai_svc_url(project_scope: AzureAIProject, token: str) -> str:
     return rai_url
 
 
-async def fetch_or_reuse_token(credential: TokenCredential, token: Optional[str] = None, workspace: Optional[str] = ML_WORKSPACE) -> str:
+async def fetch_or_reuse_token(
+    credential: TokenCredential, token: Optional[str] = None, workspace: Optional[str] = ML_WORKSPACE
+) -> str:
     """Get token. Fetch a new token if the current token is near expiry
 
        :param credential: The Azure authentication credential.
@@ -622,6 +641,7 @@ async def evaluate_with_rai_service(
     annotation_task: str = Tasks.CONTENT_HARM,
     metric_display_name=None,
     evaluator_name=None,
+    scan_session_id: Optional[str] = None,
 ) -> Dict[str, Union[str, float]]:
     """Evaluate the content safety of the response using Responsible AI service
 
@@ -629,8 +649,8 @@ async def evaluate_with_rai_service(
     :type data: dict
     :param metric_name: The evaluation metric to use.
     :type metric_name: str
-    :param project_scope: The Azure AI project, which can either be a string representing the project endpoint 
-        or an instance of AzureAIProject. It contains subscription id, resource group, and project name. 
+    :param project_scope: The Azure AI project, which can either be a string representing the project endpoint
+        or an instance of AzureAIProject. It contains subscription id, resource group, and project name.
     :type project_scope: Union[str, AzureAIProject]
     :param credential: The Azure authentication credential.
     :type credential: ~azure.core.credentials.TokenCredential
@@ -640,15 +660,23 @@ async def evaluate_with_rai_service(
     :type metric_display_name: str
     :param evaluator_name: The evaluator name to use.
     :type evaluator_name: str
+    :param scan_session_id: The scan session ID to use for the evaluation.
+    :type scan_session_id: Optional[str]
     :return: The parsed annotation result.
     :rtype: Dict[str, Union[str, float]]
     """
 
     if is_onedp_project(project_scope):
-        client = AIProjectClient(endpoint=project_scope, credential=credential)
+        client = AIProjectClient(
+            endpoint=project_scope,
+            credential=credential,
+            user_agent_policy=UserAgentPolicy(base_user_agent=UserAgentSingleton().value),
+        )
         token = await fetch_or_reuse_token(credential=credential, workspace=COG_SRV_WORKSPACE)
         await ensure_service_availability_onedp(client, token, annotation_task)
-        operation_id = await submit_request_onedp(client, data, metric_name, token, annotation_task, evaluator_name)
+        operation_id = await submit_request_onedp(
+            client, data, metric_name, token, annotation_task, evaluator_name, scan_session_id
+        )
         annotation_response = cast(List[Dict], await fetch_result_onedp(client, operation_id, token))
         result = parse_response(annotation_response, metric_name, metric_display_name)
         return result
@@ -665,6 +693,7 @@ async def evaluate_with_rai_service(
 
         return result
 
+
 def generate_payload_multimodal(content_type: str, messages, metric: str) -> Dict:
     """Generate the payload for the annotation request
     :param content_type: The type of the content representing multimodal or images.
@@ -696,6 +725,7 @@ def generate_payload_multimodal(content_type: str, messages, metric: str) -> Dic
         "AnnotationTask": task,
     }
 
+
 async def submit_multimodal_request(messages, metric: str, rai_svc_url: str, token: str) -> str:
     """Submit request to Responsible AI service for evaluation and return operation ID
     :param messages: The normalized list of messages to be entered as the "Contents" in the payload.
@@ -741,8 +771,9 @@ async def submit_multimodal_request(messages, metric: str, rai_svc_url: str, tok
     operation_id = result["location"].split("/")[-1]
     return operation_id
 
+
 async def submit_multimodal_request_onedp(client: AIProjectClient, messages, metric: str, token: str) -> str:
-    
+
     #  handle inference sdk strongly type messages
     if len(messages) > 0 and not isinstance(messages[0], dict):
         try:
@@ -758,37 +789,42 @@ async def submit_multimodal_request_onedp(client: AIProjectClient, messages, met
     ## fetch system and assistant messages from the list of messages
     filtered_messages = [message for message in messages if message["role"] != "system"]
     assistant_messages = [message for message in messages if message["role"] == "assistant"]
-    
+
     ## prepare for request
     content_type = retrieve_content_type(assistant_messages, metric)
     payload = generate_payload_multimodal(content_type, filtered_messages, metric)
     headers = get_common_headers(token)
-    
+
     response = client.evaluations.submit_annotation(payload, headers=headers)
-    
+
     result = json.loads(response)
     operation_id = result["location"].split("/")[-1]
     return operation_id
 
+
 async def evaluate_with_rai_service_multimodal(
     messages, metric_name: str, project_scope: Union[str, AzureAIProject], credential: TokenCredential
 ):
     """ "Evaluate the content safety of the response using Responsible AI service
-       :param messages: The normalized list of messages.
-       :type messages: str
-       :param metric_name: The evaluation metric to use.
-       :type metric_name: str
-       :param project_scope: The Azure AI project, which can either be a string representing the project endpoint 
-            or an instance of AzureAIProject. It contains subscription id, resource group, and project name. 
-       :type project_scope: Union[str, AzureAIProject]
-       :param credential: The Azure authentication credential.
-       :type credential: ~azure.core.credentials.TokenCredential
-       :return: The parsed annotation result.
-       :rtype: List[List[Dict]]
+    :param messages: The normalized list of messages.
+    :type messages: str
+    :param metric_name: The evaluation metric to use.
+    :type metric_name: str
+    :param project_scope: The Azure AI project, which can either be a string representing the project endpoint
+         or an instance of AzureAIProject. It contains subscription id, resource group, and project name.
+    :type project_scope: Union[str, AzureAIProject]
+    :param credential: The Azure authentication credential.
+    :type credential: ~azure.core.credentials.TokenCredential
+    :return: The parsed annotation result.
+    :rtype: List[List[Dict]]
     """
 
     if is_onedp_project(project_scope):
-        client = AIProjectClient(endpoint=project_scope, credential=credential)
+        client = AIProjectClient(
+            endpoint=project_scope,
+            credential=credential,
+            user_agent_policy=UserAgentPolicy(base_user_agent=UserAgentSingleton().value),
+        )
         token = await fetch_or_reuse_token(credential=credential, workspace=COG_SRV_WORKSPACE)
         await ensure_service_availability_onedp(client, token, Tasks.CONTENT_HARM)
         operation_id = await submit_multimodal_request_onedp(client, messages, metric_name, token)
@@ -803,4 +839,4 @@ async def evaluate_with_rai_service_multimodal(
         operation_id = await submit_multimodal_request(messages, metric_name, rai_svc_url, token)
         annotation_response = cast(List[Dict], await fetch_result(operation_id, rai_svc_url, credential, token))
         result = parse_response(annotation_response, metric_name)
-        return result
+        return result

azure/ai/evaluation/_common/raiclient/__init__.py

@@ -24,7 +24,7 @@ except ImportError:
     _patch_all = []
 from ._patch import patch_sdk as _patch_sdk
 
-# Export GeneratedRAIClient as alias of MachineLearningServicesClient for backward compatibility 
+# Export GeneratedRAIClient as alias of MachineLearningServicesClient for backward compatibility
 
 __all__ = [
     "MachineLearningServicesClient",

azure/ai/evaluation/_common/raiclient/operations/_operations.py

@@ -112,7 +112,12 @@ def build_rai_svc_get_jail_break_dataset_with_type_request(  # pylint: disable=n
 
 
 def build_rai_svc_get_attack_objectives_request(  # pylint: disable=name-too-long
-    *, risk_types: Optional[List[str]] = None, lang: Optional[str] = None, strategy: Optional[str] = None, **kwargs: Any
+    *,
+    risk_types: Optional[List[str]] = None,
+    risk_categories: Optional[List[str]] = None,
+    lang: Optional[str] = None,
+    strategy: Optional[str] = None,
+    **kwargs: Any
 ) -> HttpRequest:
     _headers = case_insensitive_dict(kwargs.pop("headers", {}) or {})
     _params = case_insensitive_dict(kwargs.pop("params", {}) or {})
@@ -127,6 +132,10 @@ def build_rai_svc_get_attack_objectives_request(  # pylint: disable=name-too-lon
     _params["api-version"] = _SERIALIZER.query("api_version", api_version, "str")
     if risk_types is not None:
         _params["riskTypes"] = [_SERIALIZER.query("risk_types", q, "str") if q is not None else "" for q in risk_types]
+    if risk_categories is not None:
+        _params["riskCategory"] = [
+            _SERIALIZER.query("risk_categories", q, "str") if q is not None else "" for q in risk_categories
+        ]
     if lang is not None:
         _params["lang"] = _SERIALIZER.query("lang", lang, "str")
     if strategy is not None:
@@ -573,6 +582,7 @@ class RAISvcOperations:
     def get_attack_objectives(
         self,
         *,
+        risk_category: str,
         risk_types: Optional[List[str]] = None,
         lang: Optional[str] = None,
         strategy: Optional[str] = None,
@@ -580,6 +590,8 @@ class RAISvcOperations:
     ) -> List[_models.AttackObjective]:
         """Get the attack objectives.
 
+        :keyword risk_category: Risk category for the attack objectives. Required.
+        :paramtype risk_category: str
         :keyword risk_types: Risk types for the attack objectives dataset. Default value is None.
         :paramtype risk_types: list[str]
         :keyword lang: The language for the attack objectives dataset, defaults to 'en'. Default value
@@ -605,6 +617,7 @@ class RAISvcOperations:
         cls: ClsType[List[_models.AttackObjective]] = kwargs.pop("cls", None)
 
         _request = build_rai_svc_get_attack_objectives_request(
+            risk_categories=[risk_category],
             risk_types=risk_types,
             lang=lang,
             strategy=strategy,

azure/ai/evaluation/_common/utils.py

@@ -13,7 +13,7 @@ from azure.storage.blob import ContainerClient
 from typing_extensions import NotRequired, Required, TypeGuard
 from azure.ai.evaluation._legacy._adapters._errors import MissingRequiredPackage
 from azure.ai.evaluation._constants import AZURE_OPENAI_TYPE, OPENAI_TYPE
-from azure.ai.evaluation._exceptions import ErrorBlame, ErrorCategory, ErrorTarget, EvaluationException
+from azure.ai.evaluation._exceptions import ErrorMessage, ErrorBlame, ErrorCategory, ErrorTarget, EvaluationException
 from azure.ai.evaluation._model_configurations import (
     AzureAIProject,
     AzureOpenAIModelConfiguration,
@@ -126,6 +126,7 @@ def construct_prompty_model_config(
 
     return prompty_model_config
 
+
 def is_onedp_project(azure_ai_project: AzureAIProject) -> bool:
     """Check if the Azure AI project is an OneDP project.
 
@@ -138,6 +139,7 @@ def is_onedp_project(azure_ai_project: AzureAIProject) -> bool:
         return True
     return False
 
+
 def validate_azure_ai_project(o: object) -> AzureAIProject:
     fields = {"subscription_id": str, "resource_group_name": str, "project_name": str}
 
@@ -291,7 +293,8 @@ def _validate_typed_dict(o: object, t: Type[T_TypedDict]) -> T_TypedDict:
 
     return cast(T_TypedDict, o)
 
-def check_score_is_valid(score: Union[str, float], min_score = 1, max_score = 5) -> bool:
+
+def check_score_is_valid(score: Union[str, float], min_score=1, max_score=5) -> bool:
     """Check if the score is valid, i.e. is convertable to number and is in the range [min_score, max_score].
 
     :param score: The score to check.
@@ -310,6 +313,7 @@ def check_score_is_valid(score: Union[str, float], min_score = 1, max_score = 5)
 
     return min_score <= numeric_score <= max_score
 
+
 def parse_quality_evaluator_reason_score(llm_output: str, valid_score_range: str = "[1-5]") -> Tuple[float, str]:
     """Parse the output of prompt-based quality evaluators that return a score and reason.
 
@@ -481,6 +485,123 @@ def validate_conversation(conversation):
             ErrorTarget.CONTENT_SAFETY_CHAT_EVALUATOR,
         )
 
+
+def _extract_text_from_content(content):
+    text = []
+    for msg in content:
+        if "text" in msg:
+            text.append(msg["text"])
+    return text
+
+
+def _get_conversation_history(query):
+    all_user_queries = []
+    cur_user_query = []
+    all_agent_responses = []
+    cur_agent_response = []
+    for msg in query:
+        if not "role" in msg:
+            continue
+        if msg["role"] == "user" and "content" in msg:
+            if cur_agent_response != []:
+                all_agent_responses.append(cur_agent_response)
+                cur_agent_response = []
+            text_in_msg = _extract_text_from_content(msg["content"])
+            if text_in_msg:
+                cur_user_query.append(text_in_msg)
+
+        if msg["role"] == "assistant" and "content" in msg:
+            if cur_user_query != []:
+                all_user_queries.append(cur_user_query)
+                cur_user_query = []
+            text_in_msg = _extract_text_from_content(msg["content"])
+            if text_in_msg:
+                cur_agent_response.append(text_in_msg)
+    if cur_user_query != []:
+        all_user_queries.append(cur_user_query)
+    if cur_agent_response != []:
+        all_agent_responses.append(cur_agent_response)
+
+    if len(all_user_queries) != len(all_agent_responses) + 1:
+        raise EvaluationException(
+            message=ErrorMessage.MALFORMED_CONVERSATION_HISTORY,
+            internal_message=ErrorMessage.MALFORMED_CONVERSATION_HISTORY,
+            target=ErrorTarget.CONVERSATION_HISTORY_PARSING,
+            category=ErrorCategory.INVALID_VALUE,
+            blame=ErrorBlame.USER_ERROR,
+        )
+
+    return {"user_queries": all_user_queries, "agent_responses": all_agent_responses}
+
+
+def _pretty_format_conversation_history(conversation_history):
+    """Formats the conversation history for better readability."""
+    formatted_history = ""
+    for i, (user_query, agent_response) in enumerate(
+        zip(conversation_history["user_queries"], conversation_history["agent_responses"] + [None])
+    ):
+        formatted_history += f"User turn {i+1}:\n"
+        for msg in user_query:
+            formatted_history += "  " + "\n  ".join(msg)
+        formatted_history += "\n\n"
+        if agent_response:
+            formatted_history += f"Agent turn {i+1}:\n"
+            for msg in agent_response:
+                formatted_history += "  " + "\n  ".join(msg)
+            formatted_history += "\n\n"
+    return formatted_history
+
+
+def reformat_conversation_history(query, logger=None):
+    """Reformats the conversation history to a more compact representation."""
+    try:
+        conversation_history = _get_conversation_history(query)
+        return _pretty_format_conversation_history(conversation_history)
+    except:
+        # If the conversation history cannot be parsed for whatever reason (e.g. the converter format changed), the original query is returned
+        # This is a fallback to ensure that the evaluation can still proceed. However the accuracy of the evaluation will be affected.
+        # From our tests the negative impact on IntentResolution is:
+        #   Higher intra model variance (0.142 vs 0.046)
+        #   Higher inter model variance (0.345 vs 0.607)
+        #   Lower percentage of mode in Likert scale (73.4% vs 75.4%)
+        #   Lower pairwise agreement between LLMs (85% vs 90% at the pass/fail level with threshold of 3)
+        if logger:
+            logger.warning(f"Conversation history could not be parsed, falling back to original query: {query}")
+        return query
+
+
+def _get_agent_response(agent_response_msgs):
+    """Extracts the text from the agent response content."""
+    agent_response_text = []
+    for msg in agent_response_msgs:
+        if "role" in msg and msg["role"] == "assistant" and "content" in msg:
+            text = _extract_text_from_content(msg["content"])
+            if text:
+                agent_response_text.extend(text)
+    return agent_response_text
+
+
+def reformat_agent_response(response, logger=None):
+    try:
+        if response is None or response == []:
+            return ""
+        agent_response = _get_agent_response(response)
+        if agent_response == []:
+            # If no message could be extracted, likely the format changed, fallback to the original response in that case
+            if logger:
+                logger.warning(
+                    f"Empty agent response extracted, likely due to input schema change. Falling back to using the original response: {response}"
+                )
+            return response
+        return "\n".join(agent_response)
+    except:
+        # If the agent response cannot be parsed for whatever reason (e.g. the converter format changed), the original response is returned
+        # This is a fallback to ensure that the evaluation can still proceed. See comments on reformat_conversation_history for more details.
+        if logger:
+            logger.warning(f"Agent response could not be parsed, falling back to original response: {response}")
+        return response
+
+
 def upload(path: str, container_client: ContainerClient, logger=None):
     """Upload files or directories to Azure Blob Storage using a container client.
 
@@ -509,7 +630,7 @@ def upload(path: str, container_client: ContainerClient, logger=None):
     local_paths = []
 
     if os.path.isdir(path):
-        for (root, _, filenames) in os.walk(path):
+        for root, _, filenames in os.walk(path):
             upload_path = ""
             if root != path:
                 rel_path = os.path.relpath(root, path)

azure/ai/evaluation/_constants.py

@@ -81,6 +81,7 @@ class _AggregationType(enum.Enum):
     SUM = "sum"
     CUSTOM = "custom"
 
+
 class TokenScope(str, enum.Enum):
     """Defines the scope of the token used to access Azure resources."""
 
@@ -114,4 +115,4 @@ BINARY_AGGREGATE_SUFFIX = "binary_aggregate"
 
 AOAI_COLUMN_NAME = "aoai"
 DEFAULT_OAI_EVAL_RUN_NAME = "AI_SDK_EVAL_RUN"
-DEFAULT_AOAI_API_VERSION = "2025-04-01-preview" # Unfortunately relying on preview version for now.
+DEFAULT_AOAI_API_VERSION = "2025-04-01-preview"  # Unfortunately relying on preview version for now.

azure/ai/evaluation/_converters/__init__.py

@@ -1,3 +1,3 @@
 # ---------------------------------------------------------
 # Copyright (c) Microsoft Corporation. All rights reserved.
-# ---------------------------------------------------------
+# ---------------------------------------------------------