Add groundedness pro eval (Azure#38063)

* Adding service based groundedness

* groundedness pro eval

* remove groundedness and fix unit tests

* run black

* change evaluate label

* Update sdk/evaluation/azure-ai-evaluation/azure/ai/evaluation/_evaluators/_service_groundedness/_service_groundedness.py

Co-authored-by: Neehar Duvvuri <[email protected]>

* Update sdk/evaluation/azure-ai-evaluation/azure/ai/evaluation/_evaluators/_service_groundedness/_service_groundedness.py

Co-authored-by: Neehar Duvvuri <[email protected]>

* comments and CL

* re record tests

* black and pylint

* comments

* nits

* analysis

* re cast

* more mypy appeasement

---------

Co-authored-by: Ankit Singhal <[email protected]>
Co-authored-by: Neehar Duvvuri <[email protected]>

Author: 3 people

sdk/evaluation/azure-ai-evaluation/CHANGELOG.md

@@ -4,6 +4,7 @@
 ## 1.0.0b5 (Unreleased)
 
 ### Features Added
+- Added `GroundednessProEvaluator`, which is a service-based evaluator for determining response groundedness.
 - Groundedness detection in Non Adversarial Simulator via query/context pairs
 ```python
 import importlib.resources as pkg_resources

sdk/evaluation/azure-ai-evaluation/assets.json

@@ -2,5 +2,5 @@
   "AssetsRepo": "Azure/azure-sdk-assets",
   "AssetsRepoPrefixPath": "python",
   "TagPrefix": "python/evaluation/azure-ai-evaluation",
-  "Tag": "python/evaluation/azure-ai-evaluation_1390701e9d"
+  "Tag": "python/evaluation/azure-ai-evaluation_5551827d25"
 }

sdk/evaluation/azure-ai-evaluation/azure/ai/evaluation/__init__.py

@@ -16,6 +16,7 @@
 from ._evaluators._fluency import FluencyEvaluator
 from ._evaluators._gleu import GleuScoreEvaluator
 from ._evaluators._groundedness import GroundednessEvaluator
+from ._evaluators._service_groundedness import GroundednessProEvaluator
 from ._evaluators._meteor import MeteorScoreEvaluator
 from ._evaluators._protected_material import ProtectedMaterialEvaluator
 from ._evaluators._qa import QAEvaluator
@@ -40,6 +41,7 @@
     "F1ScoreEvaluator",
     "FluencyEvaluator",
     "GroundednessEvaluator",
+    "GroundednessProEvaluator",
     "RelevanceEvaluator",
     "SimilarityEvaluator",
     "QAEvaluator",

sdk/evaluation/azure-ai-evaluation/azure/ai/evaluation/_common/constants.py

@@ -35,6 +35,7 @@ class Tasks:
     CONTENT_HARM = "content harm"
     PROTECTED_MATERIAL = "protected material"
     XPIA = "xpia"
+    GROUNDEDNESS = "groundedness"
 
 
 class _InternalAnnotationTasks:
@@ -56,6 +57,7 @@ class EvaluationMetrics(str, Enum, metaclass=CaseInsensitiveEnumMeta):
     SEXUAL = "sexual"
     PROTECTED_MATERIAL = "protected_material"
     XPIA = "xpia"
+    GROUNDEDNESS = "generic_groundedness"
 
 
 class _InternalEvaluationMetrics(str, Enum, metaclass=CaseInsensitiveEnumMeta):

sdk/evaluation/azure-ai-evaluation/azure/ai/evaluation/_common/rai_service.py

@@ -7,8 +7,9 @@
 import re
 import time
 from ast import literal_eval
-from typing import Any, Dict, List, Optional, Union, cast
+from typing import Dict, List, Optional, Union, cast
 from urllib.parse import urlparse
+from string import Template
 
 import jwt
 
@@ -23,7 +24,6 @@
     EvaluationMetrics,
     RAIService,
     Tasks,
-    _InternalAnnotationTasks,
     _InternalEvaluationMetrics,
 )
 from .utils import get_harm_severity_level
@@ -34,6 +34,11 @@
     version = "unknown"
 USER_AGENT = "{}/{}".format("azure-ai-evaluation", version)
 
+USER_TEXT_TEMPLATE_DICT: Dict[str, Template] = {
+    "DEFAULT": Template("<Human>{$query}</><System>{$response}</>"),
+    Tasks.GROUNDEDNESS: Template('{"question": "$query", "answer": "$response", "context": "$context"}'),
+}
+
 
 def get_common_headers(token: str) -> Dict:
     """Get common headers for the HTTP request
@@ -99,27 +104,26 @@ async def ensure_service_availability(rai_svc_url: str, token: str, capability:
         )
 
 
-def generate_payload(normalized_user_text: str, metric: str) -> Dict:
+def generate_payload(normalized_user_text: str, metric: str, annotation_task: str) -> Dict:
     """Generate the payload for the annotation request
 
     :param normalized_user_text: The normalized user text to be entered as the "UserTextList" in the payload.
     :type normalized_user_text: str
     :param metric: The evaluation metric to use. This determines the task type, and whether a "MetricList" is needed
         in the payload.
     :type metric: str
+    :param annotation_task: The annotation task to be passed to service
+    :type annotation_task: str
     :return: The payload for the annotation request.
     :rtype: Dict
     """
     include_metric = True
-    task = Tasks.CONTENT_HARM
+    task = annotation_task
     if metric == EvaluationMetrics.PROTECTED_MATERIAL:
-        task = Tasks.PROTECTED_MATERIAL
         include_metric = False
     elif metric == _InternalEvaluationMetrics.ECI:
-        task = _InternalAnnotationTasks.ECI
         include_metric = False
     elif metric == EvaluationMetrics.XPIA:
-        task = Tasks.XPIA
         include_metric = False
     return (
         {
@@ -135,25 +139,25 @@ def generate_payload(normalized_user_text: str, metric: str) -> Dict:
     )
 
 
-async def submit_request(query: str, response: str, metric: str, rai_svc_url: str, token: str) -> str:
+async def submit_request(data: dict, metric: str, rai_svc_url: str, token: str, annotation_task: str) -> str:
     """Submit request to Responsible AI service for evaluation and return operation ID
 
-    :param query: The query to evaluate.
-    :type query: str
-    :param response: The response to evaluate.
-    :type response: str
+    :param data: The data to evaluate.
+    :type data: dict
     :param metric: The evaluation metric to use.
     :type metric: str
     :param rai_svc_url: The Responsible AI service URL.
     :type rai_svc_url: str
     :param token: The Azure authentication token.
     :type token: str
+    :param annotation_task: The annotation task to use.
+    :type annotation_task: str
     :return: The operation ID.
     :rtype: str
     """
-    user_text = f"<Human>{query}</><System>{response}</>"
+    user_text = USER_TEXT_TEMPLATE_DICT.get(annotation_task, USER_TEXT_TEMPLATE_DICT["DEFAULT"]).substitute(**data)
     normalized_user_text = user_text.replace("'", '\\"')
-    payload = generate_payload(normalized_user_text, metric)
+    payload = generate_payload(normalized_user_text, metric, annotation_task=annotation_task)
 
     url = rai_svc_url + "/submitannotation"
     headers = get_common_headers(token)
@@ -164,7 +168,6 @@ async def submit_request(query: str, response: str, metric: str, rai_svc_url: st
     if http_response.status_code != 202:
         print("Fail evaluating '%s' with error message: %s" % (payload["UserTextList"], http_response.text()))
         http_response.raise_for_status()
-
     result = http_response.json()
     operation_id = result["location"].split("/")[-1]
     return operation_id
@@ -208,19 +211,28 @@ async def fetch_result(operation_id: str, rai_svc_url: str, credential: TokenCre
 
 
 def parse_response(  # pylint: disable=too-many-branches,too-many-statements
-    batch_response: List[Dict], metric_name: str
+    batch_response: List[Dict], metric_name: str, metric_display_name: Optional[str] = None
 ) -> Dict[str, Union[str, float]]:
     """Parse the annotation response from Responsible AI service for a content harm evaluation.
 
     :param batch_response: The annotation response from Responsible AI service.
     :type batch_response: List[Dict]
     :param metric_name: The evaluation metric to use.
     :type metric_name: str
+    :param metric_display_name: The evaluation metric display name to use. If unset, use the metric_name.
+    :type metric_display_name: Optional[str]
     :return: The parsed annotation result.
     :rtype: Dict[str, Union[str, float]]
     """
+    if metric_display_name is None:
+        metric_display_name = metric_name
+
     # non-numeric metrics
-    if metric_name in {EvaluationMetrics.PROTECTED_MATERIAL, _InternalEvaluationMetrics.ECI, EvaluationMetrics.XPIA}:
+    if metric_name in {
+        EvaluationMetrics.PROTECTED_MATERIAL,
+        _InternalEvaluationMetrics.ECI,
+        EvaluationMetrics.XPIA,
+    }:
         if not batch_response or len(batch_response[0]) == 0 or metric_name not in batch_response[0]:
             return {}
         response = batch_response[0][metric_name]
@@ -230,38 +242,42 @@ def parse_response(  # pylint: disable=too-many-branches,too-many-statements
         result = {}
         # Use label instead of score since these are assumed to be boolean results.
         # Use math.nan as null value since it's ignored by aggregations rather than treated as 0.
-        result[metric_name + "_label"] = parsed_response["label"] if "label" in parsed_response else math.nan
-        result[metric_name + "_reason"] = parsed_response["reasoning"] if "reasoning" in parsed_response else ""
+        result[metric_display_name + "_label"] = parsed_response["label"] if "label" in parsed_response else math.nan
+        result[metric_display_name + "_reason"] = parsed_response["reasoning"] if "reasoning" in parsed_response else ""
 
         if metric_name == EvaluationMetrics.XPIA:
             # Add "manipulated_content", "intrusion" and "information_gathering" to the result
             # if present else set them to math.nan
-            result[metric_name + "_manipulated_content"] = (
+            result[metric_display_name + "_manipulated_content"] = (
                 parsed_response["manipulated_content"] if "manipulated_content" in parsed_response else math.nan
             )
-            result[metric_name + "_intrusion"] = (
+            result[metric_display_name + "_intrusion"] = (
                 parsed_response["intrusion"] if "intrusion" in parsed_response else math.nan
             )
-            result[metric_name + "_information_gathering"] = (
+            result[metric_display_name + "_information_gathering"] = (
                 parsed_response["information_gathering"] if "information_gathering" in parsed_response else math.nan
             )
         return result
-    return _parse_content_harm_response(batch_response, metric_name)
+    return _parse_content_harm_response(batch_response, metric_name, metric_display_name)
 
 
-def _parse_content_harm_response(batch_response: List[Dict], metric_name: str) -> Dict[str, Union[str, float]]:
+def _parse_content_harm_response(
+    batch_response: List[Dict], metric_name: str, metric_display_name: Optional[str] = None
+) -> Dict[str, Union[str, float]]:
     """Parse the annotation response from Responsible AI service for a content harm evaluation.
 
     :param batch_response: The annotation response from Responsible AI service.
     :type batch_response: List[Dict]
     :param metric_name: The evaluation metric to use.
     :type metric_name: str
+    :param metric_display_name: The evaluation metric display name to use. If unset, use the metric_name.
+    :type metric_display_name: Optional[str]
     :return: The parsed annotation result.
     :rtype: Dict[str, Union[str, float]]
     """
     # Fix the metric name if it's "hate_fairness"
     # Eventually we will remove this fix once the RAI service is updated
-    key = metric_name
+    key = metric_name if metric_display_name is None else metric_display_name
     if key == EvaluationMetrics.HATE_FAIRNESS:
         key = EvaluationMetrics.HATE_UNFAIRNESS
 
@@ -283,7 +299,7 @@ def _parse_content_harm_response(batch_response: List[Dict], metric_name: str) -
 
         # get content harm metric_value
         if "label" in harm_response:
-            metric_value = harm_response["label"]
+            metric_value = float(harm_response["label"])
         elif "valid" in harm_response:
             metric_value = 0 if harm_response["valid"] else math.nan
         else:
@@ -412,33 +428,40 @@ async def fetch_or_reuse_token(credential: TokenCredential, token: Optional[str]
 
 
 async def evaluate_with_rai_service(
-    query: str, response: str, metric_name: str, project_scope: AzureAIProject, credential: TokenCredential
-) -> Dict[str, Any]:
+    data: dict,
+    metric_name: str,
+    project_scope: AzureAIProject,
+    credential: TokenCredential,
+    annotation_task: str = Tasks.CONTENT_HARM,
+    metric_display_name=None,
+) -> Dict[str, Union[str, float]]:
     """ "Evaluate the content safety of the response using Responsible AI service
 
-       :param query: The query to evaluate.
-       :type query: str
-       :param response: The response to evaluate.
-       :type response: str
+       :param data: The data to evaluate.
+       :type data: dict
        :param metric_name: The evaluation metric to use.
        :type metric_name: str
        :param project_scope: The Azure AI project scope details.
        :type project_scope: Dict
        :param credential: The Azure authentication credential.
        :type credential:
     ~azure.core.credentials.TokenCredential
+       :param annotation_task: The annotation task to use.
+       :type annotation_task: str
+       :param metric_display_name: The display name of metric to use.
+       :type metric_display_name: str
        :return: The parsed annotation result.
        :rtype: Dict[str, Union[str, float]]
     """
 
     # Get RAI service URL from discovery service and check service availability
     token = await fetch_or_reuse_token(credential)
     rai_svc_url = await get_rai_svc_url(project_scope, token)
-    await ensure_service_availability(rai_svc_url, token, Tasks.CONTENT_HARM)
+    await ensure_service_availability(rai_svc_url, token, annotation_task)
 
     # Submit annotation request and fetch result
-    operation_id = await submit_request(query, response, metric_name, rai_svc_url, token)
+    operation_id = await submit_request(data, metric_name, rai_svc_url, token, annotation_task)
     annotation_response = cast(List[Dict], await fetch_result(operation_id, rai_svc_url, credential, token))
-    result = parse_response(annotation_response, metric_name)
+    result = parse_response(annotation_response, metric_name, metric_display_name)
 
     return result

sdk/evaluation/azure-ai-evaluation/azure/ai/evaluation/_common/utils.py

@@ -24,7 +24,7 @@
 T_TypedDict = TypeVar("T_TypedDict", bound=Mapping[Any, Any])
 
 
-def get_harm_severity_level(harm_score: int) -> Union[str, float]:
+def get_harm_severity_level(harm_score: Union[float, int]) -> Union[str, float]:
     """Generate harm severity level based on harm score.
 
     :param harm_score: The harm score to be evaluated.

sdk/evaluation/azure-ai-evaluation/azure/ai/evaluation/_evaluate/_evaluate.py

@@ -36,13 +36,47 @@
 
 TClient = TypeVar("TClient", ProxyClient, CodeClient)
 
+# For metrics (aggregates) whose metric names intentionally differ from their
+# originating column name, usually because the aggregation of the original value
+# means something sufficiently different.
+# Note that content safety metrics are handled seprately.
+METRIC_COLUMN_NAME_REPLACEMENTS = {
+    "groundedness_pro_label": "groundedness_pro_passing_rate",
+}
+
 
 class __EvaluatorInfo(TypedDict):
     result: pd.DataFrame
     metrics: Dict[str, Any]
     run_summary: Dict[str, Any]
 
 
+def _aggregate_other_metrics(df: pd.DataFrame) -> Tuple[List[str], Dict[str, float]]:
+    """Identify and average various metrics that need to have the metric name be replaced,
+    instead of having the metric match the originating column name.
+    :param df: The dataframe of evaluation results.
+    :type df: ~pandas.DataFrame
+    :return: A tuple; the first element is a list of dataframe columns that were aggregated,
+        and the second element is a dictionary of resultant new metric column names and their values.
+    :rtype: Tuple[List[str], Dict[str, float]]
+    """
+    renamed_cols = []
+    metric_columns = {}
+    for col in df.columns:
+        metric_prefix = col.split(".")[0]
+        metric_name = col.split(".")[1]
+        if metric_name in METRIC_COLUMN_NAME_REPLACEMENTS:
+            renamed_cols.append(col)
+            new_col_name = metric_prefix + "." + METRIC_COLUMN_NAME_REPLACEMENTS[metric_name]
+            col_with_numeric_values = pd.to_numeric(df[col], errors="coerce")
+            metric_columns[new_col_name] = round(
+                list_sum(col_with_numeric_values) / col_with_numeric_values.count(),
+                2,
+            )
+
+    return renamed_cols, metric_columns
+
+
 # pylint: disable=line-too-long
 def _aggregate_content_safety_metrics(
     df: pd.DataFrame, evaluators: Dict[str, Callable]
@@ -146,8 +180,11 @@ def _aggregate_metrics(df: pd.DataFrame, evaluators: Dict[str, Callable]) -> Dic
     # Rename certain columns as defect rates if we know that's what their aggregates represent
     # Content safety metrics
     content_safety_cols, cs_defect_rates = _aggregate_content_safety_metrics(df, evaluators)
+    other_renamed_cols, renamed_cols = _aggregate_other_metrics(df)
     handled_columns.extend(content_safety_cols)
+    handled_columns.extend(other_renamed_cols)
     defect_rates.update(cs_defect_rates)
+    defect_rates.update(renamed_cols)
     # Label-based (true/false) metrics where 'true' means 'something is wrong'
     label_cols, label_defect_rates = _aggregate_label_defect_metrics(df)
     handled_columns.extend(label_cols)