azure-ai-evaluation 1.5.0__py3-none-any.whl → 1.6.0__py3-none-any.whl

azure/ai/evaluation/_common/utils.py

@@ -1,13 +1,15 @@
 # ---------------------------------------------------------
 # Copyright (c) Microsoft Corporation. All rights reserved.
 # ---------------------------------------------------------
-
+import os
+import posixpath
 import re
 import math
 import threading
 from typing import Any, List, Literal, Mapping, Type, TypeVar, Tuple, Union, cast, get_args, get_origin
 
 import nltk
+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
@@ -124,10 +126,25 @@ 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.
+
+    :param azure_ai_project: The scope of the Azure AI project.
+    :type azure_ai_project: ~azure.ai.evaluation.AzureAIProject
+    :return: True if the Azure AI project is an OneDP project, False otherwise.
+    :rtype: bool
+    """
+    if isinstance(azure_ai_project, str):
+        return True
+    return False
 
 def validate_azure_ai_project(o: object) -> AzureAIProject:
     fields = {"subscription_id": str, "resource_group_name": str, "project_name": str}
 
+    # TODO : Add regex check for malformed project uri
+    if is_onedp_project(o):
+        return o
+
     if not isinstance(o, dict):
         msg = "The 'azure_ai_project' parameter must be a dictionary."
         raise EvaluationException(
@@ -463,3 +480,64 @@ def validate_conversation(conversation):
             "User and assistant role expected as the only role in each message.",
             ErrorTarget.CONTENT_SAFETY_CHAT_EVALUATOR,
         )
+
+def upload(path: str, container_client: ContainerClient, logger=None):
+    """Upload files or directories to Azure Blob Storage using a container client.
+
+    This function uploads a file or all files in a directory (recursively) to Azure Blob Storage.
+    When uploading a directory, the relative path structure is preserved in the blob container.
+
+    :param path: The local path to a file or directory to upload
+    :type path: str
+    :param container_client: The Azure Blob Container client to use for uploading
+    :type container_client: azure.storage.blob.ContainerClient
+    :param logger: Optional logger for debug output, defaults to None
+    :type logger: logging.Logger, optional
+    :raises EvaluationException: If the path doesn't exist or errors occur during upload
+    """
+
+    if not os.path.isdir(path) and not os.path.isfile(path):
+        raise EvaluationException(
+            message=f"Path '{path}' is not a directory or a file",
+            internal_message=f"Path '{path}' is not a directory or a file",
+            target=ErrorTarget.RAI_CLIENT,
+            category=ErrorCategory.INVALID_VALUE,
+            blame=ErrorBlame.SYSTEM_ERROR,
+        )
+
+    remote_paths = []
+    local_paths = []
+
+    if os.path.isdir(path):
+        for (root, _, filenames) in os.walk(path):
+            upload_path = ""
+            if root != path:
+                rel_path = os.path.relpath(root, path)
+                upload_path = posixpath.join(rel_path)
+            for f in filenames:
+                remote_file_path = posixpath.join(upload_path, f)
+                remote_paths.append(remote_file_path)
+                local_file_path = os.path.join(root, f)
+                local_paths.append(local_file_path)
+
+    if os.path.isfile(path):
+        remote_paths = [os.path.basename(path)]
+        local_paths = [path]
+
+    try:
+        # Open the file in binary read mode
+        for local, remote in zip(local_paths, remote_paths):
+            with open(local, "rb") as data:
+                # Upload the file to Azure Blob Storage
+                container_client.upload_blob(data=data, name=remote)
+            if logger:
+                logger.debug(f"File '{local}' uploaded successfully")
+
+    except Exception as e:
+        raise EvaluationException(
+            message=f"Error uploading file: {e}",
+            internal_message=f"Error uploading file: {e}",
+            target=ErrorTarget.RAI_CLIENT,
+            category=ErrorCategory.UPLOAD_ERROR,
+            blame=ErrorBlame.SYSTEM_ERROR,
+        )

azure/ai/evaluation/_constants.py

@@ -62,6 +62,8 @@ class EvaluationRunProperties:
     RUN_TYPE = "runType"
     EVALUATION_RUN = "_azureml.evaluation_run"
     EVALUATION_SDK = "_azureml.evaluation_sdk_name"
+    NAME_MAP = "_azureml.evaluation_name_map"
+    NAME_MAP_LENGTH = "_azureml.evaluation_name_map_length"
 
 
 @experimental
@@ -79,6 +81,13 @@ class _AggregationType(enum.Enum):
     SUM = "sum"
     CUSTOM = "custom"
 
+class TokenScope(str, enum.Enum):
+    """Defines the scope of the token used to access Azure resources."""
+
+    DEFAULT_AZURE_MANAGEMENT = "https://management.azure.com/.default"
+    COGNITIVE_SERVICES_MANAGEMENT = "https://ai.azure.com/.default"
+    AZURE_ML = "https://ml.azure.com/.default"
+
 
 DEFAULT_EVALUATION_RESULTS_FILE_NAME = "evaluation_results.json"
 
@@ -99,3 +108,10 @@ EVALUATION_PASS_FAIL_MAPPING = {
     True: "pass",
     False: "fail",
 }
+
+DEFAULT_MAX_COMPLETION_TOKENS_REASONING_MODELS = 60000
+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.

azure/ai/evaluation/_eval_mapping.py

@@ -0,0 +1,71 @@
+# ---------------------------------------------------------
+# Copyright (c) Microsoft Corporation. All rights reserved.
+# ---------------------------------------------------------
+
+# Note: This was removed from the normal constants file due to circular import issues.
+
+# In the future, it would be nice to instead rely on the id value
+# of each eval class, but I wouldn't like to rely on those before
+# we simplify them into version-less, static values, instead of the
+# problematic registry references they currently are.
+
+# Import all evals
+from azure.ai.evaluation._evaluators._eci._eci import ECIEvaluator
+from azure.ai.evaluation import (
+    BleuScoreEvaluator,
+    CodeVulnerabilityEvaluator,
+    CoherenceEvaluator,
+    ContentSafetyEvaluator,
+    F1ScoreEvaluator,
+    FluencyEvaluator,
+    GleuScoreEvaluator,
+    GroundednessEvaluator,
+    GroundednessProEvaluator,
+    HateUnfairnessEvaluator,
+    IndirectAttackEvaluator,
+    IntentResolutionEvaluator,
+    MeteorScoreEvaluator,
+    ProtectedMaterialEvaluator,
+    QAEvaluator,
+    RelevanceEvaluator,
+    ResponseCompletenessEvaluator,
+    RetrievalEvaluator,
+    RougeScoreEvaluator,
+    SelfHarmEvaluator,
+    SexualEvaluator,
+    SimilarityEvaluator,
+    TaskAdherenceEvaluator,
+    ToolCallAccuracyEvaluator,
+    UngroundedAttributesEvaluator,
+    ViolenceEvaluator
+)
+
+EVAL_CLASS_MAP = {
+    BleuScoreEvaluator: "bleu_score",
+    CodeVulnerabilityEvaluator: "code_vulnerability",
+    CoherenceEvaluator: "coherence",
+    ContentSafetyEvaluator: "content_safety",
+    ECIEvaluator: "eci",
+    F1ScoreEvaluator: "f1_score",
+    FluencyEvaluator: "fluency",
+    GleuScoreEvaluator: "gleu_score",
+    GroundednessEvaluator: "groundedness",
+    GroundednessProEvaluator: "groundedness_pro",
+    HateUnfairnessEvaluator: "hate_unfairness",
+    IndirectAttackEvaluator: "indirect_attack",
+    IntentResolutionEvaluator: "intent_resolution",
+    MeteorScoreEvaluator: "meteor_score",
+    ProtectedMaterialEvaluator: "protected_material",
+    QAEvaluator: "qa",
+    RelevanceEvaluator: "relevance",
+    ResponseCompletenessEvaluator: "response_completeness",
+    RetrievalEvaluator: "retrieval",
+    RougeScoreEvaluator: "rouge_score",
+    SelfHarmEvaluator: "self_harm",
+    SexualEvaluator: "sexual",
+    SimilarityEvaluator: "similarity",
+    TaskAdherenceEvaluator: "task_adherence",
+    ToolCallAccuracyEvaluator: "tool_call_accuracy",
+    UngroundedAttributesEvaluator: "ungrounded_attributes",
+    ViolenceEvaluator: "violence",
+}

azure/ai/evaluation/_evaluate/_batch_run/_run_submitter_client.py

@@ -2,11 +2,12 @@
 # Copyright (c) Microsoft Corporation. All rights reserved.
 # ---------------------------------------------------------
 
+import asyncio
 import logging
 import pandas as pd
 import sys
 from collections import defaultdict
-from concurrent.futures import Future, ThreadPoolExecutor
+from concurrent.futures import Future
 from os import PathLike
 from typing import Any, Callable, Dict, Final, List, Mapping, Optional, Sequence, Union, cast
 
@@ -14,6 +15,8 @@ from .batch_clients import BatchClientRun, HasAsyncCallable
 from ..._legacy._batch_engine._run_submitter import RunSubmitter
 from ..._legacy._batch_engine._config import BatchEngineConfig
 from ..._legacy._batch_engine._run import Run
+from ..._legacy._adapters._constants import LINE_NUMBER
+from ..._legacy._common._thread_pool_executor_with_context import ThreadPoolExecutorWithContext
 
 
 LOGGER = logging.getLogger(__name__)
@@ -22,7 +25,9 @@ LOGGER = logging.getLogger(__name__)
 class RunSubmitterClient:
     def __init__(self, config: Optional[BatchEngineConfig] = None) -> None:
         self._config = config or BatchEngineConfig(LOGGER, use_async=True)
-        self._thread_pool = ThreadPoolExecutor(thread_name_prefix="evaluators_thread")
+        self._thread_pool = ThreadPoolExecutorWithContext(
+            thread_name_prefix="evaluators_thread",
+            max_workers=self._config.max_concurrency)
 
     def run(
         self,
@@ -33,30 +38,36 @@ class RunSubmitterClient:
         **kwargs: Any,
     ) -> BatchClientRun:
         if not isinstance(data, pd.DataFrame):
-            # Should never get here
             raise ValueError("Data must be a pandas DataFrame")
-        if not column_mapping:
-            raise ValueError("Column mapping must be provided")
 
-        # The column mappings are index by data to indicate they come from the data
+        # The column mappings are indexed by data to indicate they come from the data
         # input. Update the inputs so that each entry is a dictionary with a data key
         # that contains the original input data.
         inputs = [{"data": input_data} for input_data in data.to_dict(orient="records")]
 
-        # always uses async behind the scenes
+        # Pass the correct previous run to the evaluator
+        run: Optional[BatchClientRun] = kwargs.pop("run", None)
+        if run:
+            kwargs["run"] = self._get_run(run)
+
+        # Try to get async function to use
         if isinstance(flow, HasAsyncCallable):
             flow = flow._to_async()  # pylint: disable=protected-access
 
-        run_submitter = RunSubmitter(self._config)
+        # Start an event loop for async execution on a thread pool thread to separate it
+        # from the caller's thread.
+        run_submitter = RunSubmitter(self._config, self._thread_pool)
         run_future = self._thread_pool.submit(
-            run_submitter.submit,
-            dynamic_callable=flow,
-            inputs=inputs,
-            column_mapping=column_mapping,
-            name_prefix=evaluator_name,
-            created_on=kwargs.pop("created_on", None),
-            storage_creator=kwargs.pop("storage_creator", None),
-            **kwargs,
+            asyncio.run,
+            run_submitter.submit(
+                dynamic_callable=flow,
+                inputs=inputs,
+                column_mapping=column_mapping,
+                name_prefix=evaluator_name,
+                created_on=kwargs.pop("created_on", None),
+                storage_creator=kwargs.pop("storage_creator", None),
+                **kwargs,
+            )
         )
 
         return run_future
@@ -75,7 +86,10 @@ class RunSubmitterClient:
                     key = f"{prefix}.{k}"
                     data[key].append(value)
 
+        # Go from a list of dictionaries (i.e. a row view of the data) to a dictionary of lists
+        # (i.e. a column view of the data)
         _update("inputs", run.inputs)
+        _update("inputs", [{ LINE_NUMBER: i } for i in range(len(run.inputs)) ])
         _update("outputs", run.outputs)
 
         df = pd.DataFrame(data).reindex(columns=[k for k in data.keys()])

azure/ai/evaluation/_evaluate/_batch_run/eval_run_context.py

@@ -8,6 +8,10 @@ from typing import Optional, Type, Union
 from azure.ai.evaluation._legacy._adapters._constants import PF_FLOW_ENTRY_IN_TMP, PF_FLOW_META_LOAD_IN_SUBPROCESS
 from azure.ai.evaluation._legacy._adapters.utils import ClientUserAgentUtil
 from azure.ai.evaluation._legacy._adapters.tracing import inject_openai_api, recover_openai_api
+from azure.ai.evaluation._legacy._batch_engine._openai_injector import (
+    inject_openai_api as ported_inject_openai_api,
+    recover_openai_api as ported_recover_openai_api,
+)
 
 from azure.ai.evaluation._constants import (
     OTEL_EXPORTER_OTLP_TRACES_TIMEOUT,
@@ -68,6 +72,7 @@ class EvalRunContext:
 
         if isinstance(self.client, RunSubmitterClient):
             set_event_loop_policy()
+            ported_inject_openai_api()
 
     def __exit__(
         self,
@@ -92,3 +97,6 @@ class EvalRunContext:
             if self._is_otel_timeout_set_by_system:
                 os.environ.pop(OTEL_EXPORTER_OTLP_TRACES_TIMEOUT, None)
                 self._is_otel_timeout_set_by_system = False
+
+        if isinstance(self.client, RunSubmitterClient):
+            ported_recover_openai_api()

azure/ai/evaluation/_evaluate/_batch_run/proxy_client.py

@@ -58,6 +58,11 @@ class ProxyClient:  # pylint: disable=client-accepts-api-version-keyword
         if not name:
             name = f"azure_ai_evaluation_evaluators_{evaluator_name}_{datetime.now().strftime('%Y%m%d_%H%M%S_%f')}"
 
+        # Pass the correct previous run to the evaluator
+        run: Optional[BatchClientRun] = kwargs.pop("run", None)
+        if run:
+            kwargs["run"] = self.get_result(run)
+
         batch_use_async = self._should_batch_use_async(flow_to_run)
         eval_future = self._thread_pool.submit(
             self._pf_client.run,

azure/ai/evaluation/_evaluate/_batch_run/target_run_context.py

@@ -5,8 +5,15 @@ import os
 import types
 from typing import Optional, Type
 
+from azure.ai.evaluation._evaluate._batch_run.batch_clients import BatchClient
+from azure.ai.evaluation._evaluate._batch_run import RunSubmitterClient
 from azure.ai.evaluation._legacy._adapters._constants import PF_FLOW_ENTRY_IN_TMP
+from azure.ai.evaluation._legacy._batch_engine._openai_injector import (
+    inject_openai_api as ported_inject_openai_api,
+    recover_openai_api as ported_recover_openai_api,
+)
 from azure.ai.evaluation._constants import PF_DISABLE_TRACING
+from azure.ai.evaluation._evaluate._utils import set_event_loop_policy
 
 
 class TargetRunContext:
@@ -16,7 +23,8 @@ class TargetRunContext:
     :type upload_snapshot: bool
     """
 
-    def __init__(self, upload_snapshot: bool = False) -> None:
+    def __init__(self, client: BatchClient, upload_snapshot: bool = False) -> None:
+        self._client = client
         self._upload_snapshot = upload_snapshot
         self._original_cwd = os.getcwd()
 
@@ -32,6 +40,11 @@ class TargetRunContext:
 
         os.environ[PF_DISABLE_TRACING] = "true"
 
+        if isinstance(self._client, RunSubmitterClient):
+            ported_inject_openai_api()
+            # For addressing the issue of asyncio event loop closed on Windows
+            set_event_loop_policy()
+
     def __exit__(
         self,
         exc_type: Optional[Type[BaseException]],
@@ -44,3 +57,6 @@ class TargetRunContext:
             os.environ.pop(PF_FLOW_ENTRY_IN_TMP, None)
 
         os.environ.pop(PF_DISABLE_TRACING, None)
+
+        if isinstance(self._client, RunSubmitterClient):
+            ported_recover_openai_api()

azure/ai/evaluation/_evaluate/_eval_run.py

@@ -295,7 +295,7 @@ class EvalRun(contextlib.AbstractContextManager):  # pylint: disable=too-many-in
         return f"https://{self._url_base}" "/mlflow/v2.0" f"{self._get_scope()}" f"/api/2.0/mlflow/runs/log-metric"
 
     def _get_token(self) -> str:
-        return self._management_client.get_token()
+        return self._management_client.get_token().token
 
     def request_with_retry(
         self, url: str, method: str, json_dict: Dict[str, Any], headers: Optional[Dict[str, str]] = None