azure-ai-evaluation 1.11.1__py3-none-any.whl → 1.12.0__py3-none-any.whl

azure/ai/evaluation/_aoai/aoai_grader.py

@@ -1,13 +1,19 @@
 # ---------------------------------------------------------
 # Copyright (c) Microsoft Corporation. All rights reserved.
 # ---------------------------------------------------------
-from azure.ai.evaluation._model_configurations import AzureOpenAIModelConfiguration, OpenAIModelConfiguration
+from typing import TYPE_CHECKING, Any, Dict, Optional, Union
+
+from typing_extensions import TypeIs
 
-from azure.ai.evaluation._constants import DEFAULT_AOAI_API_VERSION
+from azure.ai.evaluation._common._experimental import experimental
+from azure.ai.evaluation._constants import DEFAULT_AOAI_API_VERSION, TokenScope
 from azure.ai.evaluation._exceptions import ErrorBlame, ErrorCategory, ErrorTarget, EvaluationException
+from azure.ai.evaluation._model_configurations import AzureOpenAIModelConfiguration, OpenAIModelConfiguration
 from azure.ai.evaluation._user_agent import UserAgentSingleton
-from typing import Any, Dict, Union
-from azure.ai.evaluation._common._experimental import experimental
+from azure.core.credentials import TokenCredential
+
+if TYPE_CHECKING:
+    from openai.lib.azure import AzureADTokenProvider
 
 
 @experimental
@@ -30,6 +36,8 @@ class AzureOpenAIGrader:
         to be formatted as a dictionary that matches the specifications of the sub-types of
         the TestingCriterion alias specified in (OpenAI's SDK)[https://github.com/openai/openai-python/blob/ed53107e10e6c86754866b48f8bd862659134ca8/src/openai/types/eval_create_params.py#L151].
     :type grader_config: Dict[str, Any]
+    :param credential: The credential to use to authenticate to the model. Only applicable to AzureOpenAI models.
+    :type credential: ~azure.core.credentials.TokenCredential
     :param kwargs: Additional keyword arguments to pass to the grader.
     :type kwargs: Any
 
@@ -43,10 +51,12 @@ class AzureOpenAIGrader:
         *,
         model_config: Union[AzureOpenAIModelConfiguration, OpenAIModelConfiguration],
         grader_config: Dict[str, Any],
+        credential: Optional[TokenCredential] = None,
         **kwargs: Any,
     ):
         self._model_config = model_config
         self._grader_config = grader_config
+        self._credential = credential
 
         if kwargs.get("validate", True):
             self._validate_model_config()
@@ -54,20 +64,39 @@ class AzureOpenAIGrader:
 
     def _validate_model_config(self) -> None:
         """Validate the model configuration that this grader wrapper is using."""
-        if "api_key" not in self._model_config or not self._model_config.get("api_key"):
-            msg = f"{type(self).__name__}: Requires an api_key in the supplied model_config."
-            raise EvaluationException(
-                message=msg,
-                blame=ErrorBlame.USER_ERROR,
-                category=ErrorCategory.INVALID_VALUE,
-                target=ErrorTarget.AOAI_GRADER,
-            )
+        msg = None
+        if self._is_azure_model_config(self._model_config):
+            if not any(auth for auth in (self._model_config.get("api_key"), self._credential)):
+                msg = (
+                    f"{type(self).__name__}: Requires an api_key in the supplied model_config, "
+                    + "or providing a credential to the grader's __init__ method. "
+                )
+
+        else:
+            if "api_key" not in self._model_config or not self._model_config.get("api_key"):
+                msg = f"{type(self).__name__}: Requires an api_key in the supplied model_config."
+
+        if msg is None:
+            return
+
+        raise EvaluationException(
+            message=msg,
+            blame=ErrorBlame.USER_ERROR,
+            category=ErrorCategory.INVALID_VALUE,
+            target=ErrorTarget.AOAI_GRADER,
+        )
 
     def _validate_grader_config(self) -> None:
         """Validate the grader configuration that this grader wrapper is using."""
 
         return
 
+    @staticmethod
+    def _is_azure_model_config(
+        model_config: Union[AzureOpenAIModelConfiguration, OpenAIModelConfiguration],
+    ) -> TypeIs[AzureOpenAIModelConfiguration]:
+        return "azure_endpoint" in model_config
+
     def get_client(self) -> Any:
         """Construct an appropriate OpenAI client using this grader's model configuration.
         Returns a slightly different client depending on whether or not this grader's model
@@ -77,23 +106,38 @@ class AzureOpenAIGrader:
         :rtype: [~openai.OpenAI, ~openai.AzureOpenAI]
         """
         default_headers = {"User-Agent": UserAgentSingleton().value}
-        if "azure_endpoint" in self._model_config:
+        model_config: Union[AzureOpenAIModelConfiguration, OpenAIModelConfiguration] = self._model_config
+        api_key: Optional[str] = model_config.get("api_key")
+
+        if self._is_azure_model_config(model_config):
             from openai import AzureOpenAI
 
             # TODO set default values?
             return AzureOpenAI(
-                azure_endpoint=self._model_config["azure_endpoint"],
-                api_key=self._model_config.get("api_key", None),  # Default-style access to appease linters.
+                azure_endpoint=model_config["azure_endpoint"],
+                api_key=api_key,  # Default-style access to appease linters.
                 api_version=DEFAULT_AOAI_API_VERSION,  # Force a known working version
-                azure_deployment=self._model_config.get("azure_deployment", ""),
+                azure_deployment=model_config.get("azure_deployment", ""),
+                azure_ad_token_provider=self._get_token_provider(self._credential) if not api_key else None,
                 default_headers=default_headers,
             )
         from openai import OpenAI
 
         # TODO add default values for base_url and organization?
         return OpenAI(
-            api_key=self._model_config["api_key"],
-            base_url=self._model_config.get("base_url", ""),
-            organization=self._model_config.get("organization", ""),
+            api_key=api_key,
+            base_url=model_config.get("base_url", ""),
+            organization=model_config.get("organization", ""),
             default_headers=default_headers,
         )
+
+    @staticmethod
+    def _get_token_provider(cred: TokenCredential) -> "AzureADTokenProvider":
+        """Get the token provider the AzureOpenAI client.
+
+        :param TokenCredential cred: The Azure authentication credential.
+        :return: The token provider if a credential is provided, otherwise None.
+        :rtype: openai.lib.azure.AzureADTokenProvider
+        """
+
+        return lambda: cred.get_token(TokenScope.COGNITIVE_SERVICES_MANAGEMENT).token

azure/ai/evaluation/_aoai/label_grader.py

@@ -1,11 +1,13 @@
 # ---------------------------------------------------------
 # Copyright (c) Microsoft Corporation. All rights reserved.
 # ---------------------------------------------------------
-from typing import Any, Dict, Union, List
+from typing import Any, Dict, List, Optional, Union
 
-from azure.ai.evaluation._model_configurations import AzureOpenAIModelConfiguration, OpenAIModelConfiguration
 from openai.types.graders import LabelModelGrader
+
 from azure.ai.evaluation._common._experimental import experimental
+from azure.ai.evaluation._model_configurations import AzureOpenAIModelConfiguration, OpenAIModelConfiguration
+from azure.core.credentials import TokenCredential
 
 from .aoai_grader import AzureOpenAIGrader
 
@@ -37,6 +39,8 @@ class AzureOpenAILabelGrader(AzureOpenAIGrader):
     :type name: str
     :param passing_labels: The labels that indicate a passing result. Must be a subset of labels.
     :type passing_labels: List[str]
+    :param credential: The credential to use to authenticate to the model. Only applicable to AzureOpenAI models.
+    :type credential: ~azure.core.credentials.TokenCredential
     :param kwargs: Additional keyword arguments to pass to the grader.
     :type kwargs: Any
 
@@ -54,6 +58,7 @@ class AzureOpenAILabelGrader(AzureOpenAIGrader):
         model: str,
         name: str,
         passing_labels: List[str],
+        credential: Optional[TokenCredential] = None,
         **kwargs: Any
     ):
         grader = LabelModelGrader(
@@ -64,4 +69,4 @@ class AzureOpenAILabelGrader(AzureOpenAIGrader):
             passing_labels=passing_labels,
             type="label_model",
         )
-        super().__init__(model_config=model_config, grader_config=grader, **kwargs)
+        super().__init__(model_config=model_config, grader_config=grader, credential=credential, **kwargs)

azure/ai/evaluation/_aoai/python_grader.py

@@ -1,11 +1,13 @@
 # ---------------------------------------------------------
 # Copyright (c) Microsoft Corporation. All rights reserved.
 # ---------------------------------------------------------
-from typing import Any, Dict, Union, Optional
+from typing import Any, Dict, Optional, Union
 
-from azure.ai.evaluation._model_configurations import AzureOpenAIModelConfiguration, OpenAIModelConfiguration
 from openai.types.graders import PythonGrader
+
 from azure.ai.evaluation._common._experimental import experimental
+from azure.ai.evaluation._model_configurations import AzureOpenAIModelConfiguration, OpenAIModelConfiguration
+from azure.core.credentials import TokenCredential
 
 from .aoai_grader import AzureOpenAIGrader
 
@@ -39,6 +41,8 @@ class AzureOpenAIPythonGrader(AzureOpenAIGrader):
     :param source: Python source code containing the grade function.
         Must define: def grade(sample: dict, item: dict) -> float
     :type source: str
+    :param credential: The credential to use to authenticate to the model. Only applicable to AzureOpenAI models.
+    :type credential: ~azure.core.credentials.TokenCredential
     :param kwargs: Additional keyword arguments to pass to the grader.
     :type kwargs: Any
 
@@ -63,6 +67,7 @@ class AzureOpenAIPythonGrader(AzureOpenAIGrader):
         image_tag: str,
         pass_threshold: float,
         source: str,
+        credential: Optional[TokenCredential] = None,
         **kwargs: Any,
     ):
         # Validate pass_threshold
@@ -81,4 +86,4 @@ class AzureOpenAIPythonGrader(AzureOpenAIGrader):
             type="python",
         )
 
-        super().__init__(model_config=model_config, grader_config=grader, **kwargs)
+        super().__init__(model_config=model_config, grader_config=grader, credential=credential, **kwargs)

azure/ai/evaluation/_aoai/score_model_grader.py

@@ -1,11 +1,13 @@
 # ---------------------------------------------------------
 # Copyright (c) Microsoft Corporation. All rights reserved.
 # ---------------------------------------------------------
-from typing import Any, Dict, Union, List, Optional
+from typing import Any, Dict, List, Optional, Union
 
-from azure.ai.evaluation._model_configurations import AzureOpenAIModelConfiguration, OpenAIModelConfiguration
 from openai.types.graders import ScoreModelGrader
+
 from azure.ai.evaluation._common._experimental import experimental
+from azure.ai.evaluation._model_configurations import AzureOpenAIModelConfiguration, OpenAIModelConfiguration
+from azure.core.credentials import TokenCredential
 
 from .aoai_grader import AzureOpenAIGrader
 
@@ -43,6 +45,8 @@ class AzureOpenAIScoreModelGrader(AzureOpenAIGrader):
     :type pass_threshold: Optional[float]
     :param sampling_params: The sampling parameters for the model.
     :type sampling_params: Optional[Dict[str, Any]]
+    :param credential: The credential to use to authenticate to the model. Only applicable to AzureOpenAI models.
+    :type credential: ~azure.core.credentials.TokenCredential
     :param kwargs: Additional keyword arguments to pass to the grader.
     :type kwargs: Any
     """
@@ -59,6 +63,7 @@ class AzureOpenAIScoreModelGrader(AzureOpenAIGrader):
         range: Optional[List[float]] = None,
         pass_threshold: Optional[float] = None,
         sampling_params: Optional[Dict[str, Any]] = None,
+        credential: Optional[TokenCredential] = None,
         **kwargs: Any,
     ):
         # Validate range and pass_threshold
@@ -88,4 +93,4 @@ class AzureOpenAIScoreModelGrader(AzureOpenAIGrader):
 
         grader = ScoreModelGrader(**grader_kwargs)
 
-        super().__init__(model_config=model_config, grader_config=grader, **kwargs)
+        super().__init__(model_config=model_config, grader_config=grader, credential=credential, **kwargs)

azure/ai/evaluation/_aoai/string_check_grader.py

@@ -1,12 +1,14 @@
 # ---------------------------------------------------------
 # Copyright (c) Microsoft Corporation. All rights reserved.
 # ---------------------------------------------------------
-from typing import Any, Dict, Union
-from typing_extensions import Literal
+from typing import Any, Dict, Optional, Union
 
-from azure.ai.evaluation._model_configurations import AzureOpenAIModelConfiguration, OpenAIModelConfiguration
 from openai.types.graders import StringCheckGrader
+from typing_extensions import Literal
+
 from azure.ai.evaluation._common._experimental import experimental
+from azure.ai.evaluation._model_configurations import AzureOpenAIModelConfiguration, OpenAIModelConfiguration
+from azure.core.credentials import TokenCredential
 
 from .aoai_grader import AzureOpenAIGrader
 
@@ -33,6 +35,8 @@ class AzureOpenAIStringCheckGrader(AzureOpenAIGrader):
     :type operation: Literal["eq", "ne", "like", "ilike"]
     :param reference: The reference text. This may include template strings.
     :type reference: str
+    :param credential: The credential to use to authenticate to the model. Only applicable to AzureOpenAI models.
+    :type credential: ~azure.core.credentials.TokenCredential
     :param kwargs: Additional keyword arguments to pass to the grader.
     :type kwargs: Any
 
@@ -54,6 +58,7 @@ class AzureOpenAIStringCheckGrader(AzureOpenAIGrader):
             "ilike",
         ],
         reference: str,
+        credential: Optional[TokenCredential] = None,
         **kwargs: Any
     ):
         grader = StringCheckGrader(
@@ -63,4 +68,4 @@ class AzureOpenAIStringCheckGrader(AzureOpenAIGrader):
             reference=reference,
             type="string_check",
         )
-        super().__init__(model_config=model_config, grader_config=grader, **kwargs)
+        super().__init__(model_config=model_config, grader_config=grader, credential=credential, **kwargs)

azure/ai/evaluation/_aoai/text_similarity_grader.py

@@ -1,12 +1,14 @@
 # ---------------------------------------------------------
 # Copyright (c) Microsoft Corporation. All rights reserved.
 # ---------------------------------------------------------
-from typing import Any, Dict, Union
-from typing_extensions import Literal
+from typing import Any, Dict, Optional, Union
 
-from azure.ai.evaluation._model_configurations import AzureOpenAIModelConfiguration, OpenAIModelConfiguration
 from openai.types.graders import TextSimilarityGrader
+from typing_extensions import Literal
+
 from azure.ai.evaluation._common._experimental import experimental
+from azure.ai.evaluation._model_configurations import AzureOpenAIModelConfiguration, OpenAIModelConfiguration
+from azure.core.credentials import TokenCredential
 
 from .aoai_grader import AzureOpenAIGrader
 
@@ -47,6 +49,8 @@ class AzureOpenAITextSimilarityGrader(AzureOpenAIGrader):
     :type reference: str
     :param name: The name of the grader.
     :type name: str
+    :param credential: The credential to use to authenticate to the model. Only applicable to AzureOpenAI models.
+    :type credential: ~azure.core.credentials.TokenCredential
     :param kwargs: Additional keyword arguments to pass to the grader.
     :type kwargs: Any
 
@@ -76,6 +80,7 @@ class AzureOpenAITextSimilarityGrader(AzureOpenAIGrader):
         pass_threshold: float,
         reference: str,
         name: str,
+        credential: Optional[TokenCredential] = None,
         **kwargs: Any
     ):
         grader = TextSimilarityGrader(
@@ -86,4 +91,4 @@ class AzureOpenAITextSimilarityGrader(AzureOpenAIGrader):
             reference=reference,
             type="text_similarity",
         )
-        super().__init__(model_config=model_config, grader_config=grader, **kwargs)
+        super().__init__(model_config=model_config, grader_config=grader, credential=credential, **kwargs)

azure/ai/evaluation/_eval_mapping.py

@@ -11,6 +11,7 @@
 
 # Import all evals
 from azure.ai.evaluation._evaluators._eci._eci import ECIEvaluator
+from azure.ai.evaluation._evaluators._task_success import TaskSuccessEvaluator
 from azure.ai.evaluation import (
     BleuScoreEvaluator,
     CodeVulnerabilityEvaluator,
@@ -67,6 +68,7 @@ EVAL_CLASS_MAP = {
     SexualEvaluator: "sexual",
     SimilarityEvaluator: "similarity",
     TaskAdherenceEvaluator: "task_adherence",
+    TaskSuccessEvaluator: "task_success",
     ToolCallAccuracyEvaluator: "tool_call_accuracy",
     UngroundedAttributesEvaluator: "ungrounded_attributes",
     ViolenceEvaluator: "violence",

azure/ai/evaluation/_evaluate/_evaluate.py

@@ -9,7 +9,7 @@ import os
 import re
 import tempfile
 import json
-from typing import Any, Callable, Dict, List, Literal, Optional, Set, Tuple, TypedDict, Union, cast
+from typing import Any, Callable, Dict, Iterable, Iterator, List, Literal, Optional, Set, Tuple, TypedDict, Union, cast
 
 from openai import OpenAI, AzureOpenAI
 from azure.ai.evaluation._legacy._adapters._constants import LINE_NUMBER
@@ -18,6 +18,7 @@ import pandas as pd
 
 from azure.ai.evaluation._common.math import list_mean_nan_safe, apply_transform_nan_safe
 from azure.ai.evaluation._common.utils import validate_azure_ai_project, is_onedp_project
+from azure.ai.evaluation._evaluators._common._base_eval import EvaluatorBase
 from azure.ai.evaluation._exceptions import ErrorBlame, ErrorCategory, ErrorTarget, EvaluationException
 
 from azure.ai.evaluation._aoai.aoai_grader import AzureOpenAIGrader
@@ -317,6 +318,9 @@ def _aggregate_metrics(df: pd.DataFrame, evaluators: Dict[str, Callable]) -> Dic
     # For rest of metrics, we will calculate mean
     df.drop(columns=handled_columns, inplace=True)
 
+    # Convert "not applicable" strings to None to allow proper numeric aggregation
+    df = df.replace(EvaluatorBase._NOT_APPLICABLE_RESULT, None)
+
     # NOTE: nan/None values don't count as as booleans, so boolean columns with
     # nan/None values won't have a mean produced from them.
     # This is different from label-based known evaluators, which have special handling.
@@ -1131,11 +1135,36 @@ def _preprocess_data(
     # via target mapping.
     # If both the data and the output dictionary of the target function
     # have the same column, then the target function value is used.
+    # NEW: flatten nested object columns (e.g., 'item') so we can map leaf values automatically.
+    # Ensure the data does not contain top-level 'conversation' or 'messages' columns (which indicate chat/conversation data)
+    if input_data_df is not None:
+        if "conversation" in input_data_df.columns or "messages" in input_data_df.columns:
+            # No action is taken when 'conversation' or 'messages' columns are present,
+            # as these indicate chat/conversation data which should not be flattened or mapped by default.
+            pass
+        else:
+            input_data_df = _flatten_object_columns_for_default_mapping(input_data_df)
+
+    # Build default mapping for leaves:
     if input_data_df is not None:
+        # First, map flattened nested columns (those containing a dot) to leaf names.
+        for col in input_data_df.columns:
+            # Skip target output columns
+            if col.startswith(Prefixes.TSG_OUTPUTS):
+                continue
+            # Skip root container columns (no dot) here; they'll be handled below if truly primitive.
+            if "." in col:
+                leaf_name = col.split(".")[-1]
+                if leaf_name not in column_mapping["default"]:
+                    column_mapping["default"][leaf_name] = f"${{data.{col}}}"
+
+        # Then, handle remaining top-level primitive columns (original logic).
         for col in input_data_df.columns:
-            # Ignore columns added by target mapping. These are formatted as "__outputs.<column_name>"
-            # Also ignore columns that are already in config, since they've been covered by target mapping.
-            if not col.startswith(Prefixes.TSG_OUTPUTS) and col not in column_mapping["default"].keys():
+            if (
+                not col.startswith(Prefixes.TSG_OUTPUTS)
+                and col not in column_mapping["default"].keys()
+                and "." not in col  # only pure top-level primitives
+            ):
                 column_mapping["default"][col] = f"${{data.{col}}}"
 
     return __ValidatedData(
@@ -1149,6 +1178,79 @@ def _preprocess_data(
     )
 
 
+def _flatten_object_columns_for_default_mapping(
+    df: pd.DataFrame, root_prefixes: Optional[Iterable[str]] = None
+) -> pd.DataFrame:
+    """Flatten nested dictionary-valued columns into dotted leaf columns.
+
+    For any column whose cells (in at least one row) are ``dict`` objects, this utility discovers all
+    leaf paths (recursively descending only through ``dict`` nodes) and materializes new DataFrame
+    columns named ``"<original_col>.<nested.path.leaf>"`` for every unique leaf encountered across
+    all rows. A *leaf* is defined as any value that is **not** a ``dict`` (lists / primitives / ``None``
+    are all treated as leaves). Existing columns are never overwritten (idempotent behavior).
+
+    Example
+        If a column ``item`` contains objects like ``{"a": {"b": 1, "c": 2}}`` a pair of new
+        columns ``item.a.b`` and ``item.a.c`` will be added with the corresponding scalar values.
+
+    :param df: Input DataFrame to flatten in place.
+    :type df: ~pandas.DataFrame
+    :param root_prefixes: Optional iterable restricting which top-level columns are considered
+        for flattening. If ``None``, all columns containing at least one ``dict`` value are processed.
+    :type root_prefixes: Optional[Iterable[str]]
+    :return: The same DataFrame instance (returned for convenient chaining).
+    :rtype: ~pandas.DataFrame
+    """
+    candidate_cols = []
+    if root_prefixes is not None:
+        candidate_cols = [c for c in root_prefixes if c in df.columns]
+    else:
+        # pick columns where at least one non-null value is a dict
+        for c in df.columns:
+            series = df[c]
+            if series.map(lambda v: isinstance(v, dict)).any():
+                candidate_cols.append(c)
+
+    def _extract_leaves(obj: Any, prefix: str) -> Iterator[Tuple[str, Any]]:
+        if isinstance(obj, dict):
+            for k, v in obj.items():
+                new_prefix = f"{prefix}.{k}" if prefix else k
+                if isinstance(v, dict):
+                    yield from _extract_leaves(v, new_prefix)
+                else:
+                    # treat list / primitive / None as leaf
+                    yield new_prefix, v
+
+    for root_col in candidate_cols:
+        # Build a union of leaf paths across rows to ensure consistent columns
+        leaf_paths: Set[str] = set()
+        for val in df[root_col]:
+            if isinstance(val, dict):
+                for path, _ in _extract_leaves(val, root_col):
+                    leaf_paths.add(path)
+
+        if not leaf_paths:
+            continue
+
+        # Create each flattened column if absent
+        for path in leaf_paths:
+            if path in df.columns:
+                continue  # already present
+            relative_keys = path[len(root_col) + 1 :].split(".") if len(path) > len(root_col) else []
+
+            def getter(root_val: Any) -> Any:
+                cur = root_val
+                for rk in relative_keys:
+                    if not isinstance(cur, dict):
+                        return None
+                    cur = cur.get(rk, None)
+                return cur
+
+            df[path] = df[root_col].map(lambda rv: getter(rv) if isinstance(rv, dict) else None)
+
+    return df
+
+
 def _run_callable_evaluators(
     validated_data: __ValidatedData,
     fail_on_evaluator_errors: bool = False,