ibm-watsonx-gov 1.3.3__cp313-cp313-win_amd64.whl

ibm_watsonx_gov/metrics/faithfulness/faithfulness_metric.py

@@ -0,0 +1,254 @@
+# ----------------------------------------------------------------------------------------------------
+# IBM Confidential
+# Licensed Materials - Property of IBM
+# 5737-H76, 5900-A3Q
+# © Copyright IBM Corp. 2025  All Rights Reserved.
+# US Government Users Restricted Rights - Use, duplication or disclosure restricted by
+# GSA ADPSchedule Contract with IBM Corp.
+# ----------------------------------------------------------------------------------------------------
+
+from typing import Annotated, Literal
+
+import pandas as pd
+from lazy_imports import LazyModule, load
+from pydantic import Field, model_validator
+from typing_extensions import Self
+
+from ibm_watsonx_gov.config import AgenticAIConfiguration, GenAIConfiguration
+from ibm_watsonx_gov.entities.base_classes import Error
+from ibm_watsonx_gov.entities.enums import MetricGroup, TaskType
+from ibm_watsonx_gov.entities.evaluation_result import (AggregateMetricResult,
+                                                        RecordMetricResult)
+from ibm_watsonx_gov.entities.llm_judge import LLMJudge
+from ibm_watsonx_gov.entities.metric import GenAIMetric
+from ibm_watsonx_gov.entities.metric_threshold import MetricThreshold
+from ibm_watsonx_gov.providers.detectors_provider import DetectorsProvider
+from ibm_watsonx_gov.utils.async_util import run_in_event_loop
+from ibm_watsonx_gov.utils.gov_sdk_logger import GovSDKLogger
+from ibm_watsonx_gov.utils.validation_util import (validate_context,
+                                                   validate_input,
+                                                   validate_llm_as_judge,
+                                                   validate_output,
+                                                   validate_small_model_method,
+                                                   validate_unitxt_method)
+
+unitxt_provider = LazyModule(
+    "from ibm_watsonx_gov.providers.unitxt_provider import UnitxtProvider",
+    name="lazy_unitxt_provider"
+)
+load(unitxt_provider)
+UnitxtProvider = unitxt_provider.UnitxtProvider
+
+logger = GovSDKLogger.get_logger(__name__)
+FAITHFULNESS = "faithfulness"
+
+unitxt_methods = [
+    "token_k_precision",
+    "sentence_bert_mini_lm",
+    "llm_as_judge",
+    "granite_guardian",
+    "faithfulness_model"
+]
+
+
+class FaithfulnessMetric(GenAIMetric):
+    """
+    Defines the Faithfulness metric class.
+
+    The faithfulness metrics can be computed using the below methods:
+
+    1. token_k_precision (default)
+    2. sentence_bert_mini_lm
+    3. llm_as_judge
+    4. granite_guardian
+    5. faithfulness_model
+
+    Examples:
+        1. Create Faithfulness metric with default parameters.
+            .. code-block:: python
+
+                metric = FaithfulnessMetric()
+                result = MetricsEvaluator().evaluate(data={"input_text": "...", "context": "...", "generated_text": "..."}, 
+                                                    metrics=[metric])
+                # A list of contexts can also be passed as shown below
+                result = MetricsEvaluator().evaluate(data={"input_text": "...", "context": ["...", "..."], "generated_text": "..."}, 
+                                                    metrics=[metric])
+
+        2. Create Faithfulness metric with a custom threshold and method.
+            .. code-block:: python
+
+                thresholds  = [MetricThreshold(type="lower_limit", value=0.5)]
+                method = "sentence_bert_mini_lm"
+                metric = FaithfulnessMetric(method=method, thresholds=thresholds)
+
+        3. Create Faithfulness metric with llm_as_judge method.
+            .. code-block:: python
+
+                # Define LLM Judge using watsonx.ai
+                # To use other frameworks and models as llm_judge, see :module:`ibm_watsonx_gov.entities.foundation_model`
+                llm_judge = LLMJudge(model=WxAIFoundationModel(
+                                            model_id="ibm/granite-3-3-8b-instruct",
+                                            project_id="<PROJECT_ID>"
+                                    ))
+                metric = FaithfulnessMetric(llm_judge=llm_judge)
+
+        4. Create Faithfulness metric with granite_guardian method.
+            .. code-block:: python
+
+                metric = FaithfulnessMetric(method="granite_guardian")
+
+        5. Create Faithfulness metric with faithfulness_model method. Currently available only in On-Prem version.
+            .. code-block:: python
+
+                metric = FaithfulnessMetric(method="faithfulness_model")
+    """
+    name: Annotated[Literal["faithfulness"],
+                    Field(title="Name",
+                          description="The faithfulness metric name.",
+                          default=FAITHFULNESS, frozen=True)]
+    display_name: Annotated[Literal["Faithfulness"],
+                            Field(title="Display Name",
+                                  description="The faithfulness metric display name.",
+                                  default="Faithfulness", frozen=True)]
+    tasks: Annotated[list[TaskType],
+                     Field(title="Tasks",
+                           description="The list of supported tasks.",
+                           default=[TaskType.RAG])]
+    thresholds: Annotated[list[MetricThreshold],
+                          Field(title="Thresholds",
+                                description="The metric thresholds.",
+                                default=[MetricThreshold(type="lower_limit", value=0.7)])]
+    method: Annotated[Literal["token_k_precision", "sentence_bert_mini_lm", "llm_as_judge", "granite_guardian", "faithfulness_model"],
+                      Field(title="Method",
+                            description="The method used to compute the metric. This field is optional and when `llm_judge` is provided, the method would be set to `llm_as_judge`. The `faithfulness_model` method is currently available only in On-Prem version.",
+                            default="token_k_precision")]
+    group: Annotated[MetricGroup,
+                     Field(title="Group",
+                           description="The metric group.",
+                           default=MetricGroup.ANSWER_QUALITY, frozen=True)]
+    llm_judge: Annotated[LLMJudge | None,
+                         Field(title="LLM Judge",
+                               description="The LLM judge used to compute the metric.",
+                               default=None)]
+
+    @model_validator(mode="after")
+    def set_llm_judge_default_method(self) -> Self:
+        # If llm_judge is set, set the method to llm_as_judge
+        if self.llm_judge:
+            self.method = "llm_as_judge"
+        return self
+
+    def evaluate(self,
+                 data: pd.DataFrame,
+                 configuration: GenAIConfiguration | AgenticAIConfiguration,
+                 **kwargs) -> AggregateMetricResult:
+        # If ran in sync mode, block until it is done
+        return run_in_event_loop(
+            self.evaluate_async,
+            data=data,
+            configuration=configuration,
+            **kwargs,
+        )
+
+    def __is_supported(self, **kwargs):
+        # Currently supported only in CPD and ypqa
+        return kwargs.get(
+            "api_client").credentials.region == "ypqa" or kwargs.get("api_client").is_cpd
+
+    async def evaluate_async(self,
+                             data: pd.DataFrame,
+                             configuration: GenAIConfiguration | AgenticAIConfiguration,
+                             **kwargs) -> AggregateMetricResult:
+        data_cols = data.columns.to_list()
+
+        try:
+            validate_input(data_cols, configuration)
+            validate_output(data_cols, configuration)
+            validate_context(data_cols, configuration)
+            validate_unitxt_method(self.name, self.method, unitxt_methods)
+            validate_llm_as_judge(self.name, self.method,
+                                  self.llm_judge, configuration.llm_judge)
+            validate_small_model_method(
+                self.name, self.method, self.__is_supported(**kwargs), unitxt_methods)
+        except ValueError as ve:
+            if kwargs.get("ignore_validation_errors"):
+                message = f"Skipping '{self.name}' computation because the validation failed. Details: {str(ve)}"
+                logger.warning(message)
+                return
+            raise ve
+
+        # Separate the data into a dataframe with no None values and a dataframe with None values
+        input_output_has_none = data[configuration.input_fields +
+                                     configuration.output_fields].isna().any(axis=1)
+        all_contexts_none = data[configuration.context_fields].isna().all(
+            axis=1)
+        mask_has_none = input_output_has_none | all_contexts_none
+        df_with_none = data[mask_has_none]
+        df_without_none = data[mask_has_none == False]
+
+        # Compute the metrics only for the dataframe with no None values
+        aggregated_metric_result = None
+        if not df_without_none.empty:
+            if self.method in ["granite_guardian", "faithfulness_model"]:
+                kwargs["detector_params"] = {
+                    "method": self.method, "threshold": 0.001}
+                provider = DetectorsProvider(configuration=configuration,
+                                             metric_name=self.name,
+                                             metric_display_name=self.display_name,
+                                             metric_method=self.method,
+                                             metric_group=self.group,
+                                             thresholds=self.thresholds,
+                                             **kwargs)
+            else:
+                provider = UnitxtProvider(configuration=configuration,
+                                          metric_name=self.name,
+                                          metric_display_name=self.display_name,
+                                          metric_method=self.method,
+                                          metric_prefix="metrics.rag.external_rag",
+                                          metric_group=self.group,
+                                          llm_judge=self.llm_judge,
+                                          thresholds=self.thresholds,
+                                          **kwargs)
+
+            aggregated_metric_result = await provider.evaluate_async(data=df_without_none)
+
+        # Update the metric result with record level metrics results for the records with missing values
+        if not df_with_none.empty:
+            # Create None results for records with missing values
+            none_results = []
+            for _, row in df_with_none.iterrows():
+                record_result = RecordMetricResult(
+                    name=self.name,
+                    display_name=self.display_name,
+                    method=self.method,
+                    group=self.group,
+                    value=None,
+                    record_id=row[configuration.record_id_field],
+                    thresholds=self.thresholds,
+                    errors=[Error(
+                        code="BAD_REQUEST", message_en="The value of required fields input, output or context is None.")]
+                )
+                none_results.append(record_result)
+
+            # Merge the results
+            if aggregated_metric_result:
+                all_record_results = aggregated_metric_result.record_level_metrics + none_results
+                aggregated_metric_result.record_level_metrics = all_record_results
+                aggregated_metric_result.total_records = len(
+                    all_record_results)
+            else:
+                aggregated_metric_result = AggregateMetricResult(
+                    name=self.name,
+                    display_name=self.display_name,
+                    method=self.method,
+                    group=self.group,
+                    value=None,
+                    total_records=len(none_results),
+                    record_level_metrics=none_results,
+                    min=None,
+                    max=None,
+                    mean=None,
+                    thresholds=self.thresholds
+                )
+
+        return aggregated_metric_result

ibm_watsonx_gov/metrics/hap/__init__.py

@@ -0,0 +1,16 @@
+# ----------------------------------------------------------------------------------------------------
+# IBM Confidential
+# Licensed Materials - Property of IBM
+# 5737-H76, 5900-A3Q
+# © Copyright IBM Corp. 2025  All Rights Reserved.
+# US Government Users Restricted Rights - Use, duplication or disclosure restricted by
+# GSA ADPSchedule Contract with IBM Corp.
+# ----------------------------------------------------------------------------------------------------
+
+from ibm_watsonx_gov.metrics.hap.hap_metric import HAPMetric
+from ibm_watsonx_gov.metrics.hap.input_hap_metric import InputHAPMetric
+from ibm_watsonx_gov.metrics.hap.output_hap_metric import OutputHAPMetric
+
+__all__ = ["HAPMetric", "InputHAPMetric", "OutputHAPMetric"]
+
+# Made with Bob

ibm_watsonx_gov/metrics/hap/hap_decorator.py

@@ -0,0 +1,58 @@
+# ----------------------------------------------------------------------------------------------------
+# IBM Confidential
+# Licensed Materials - Property of IBM
+# 5737-H76, 5900-A3Q
+# © Copyright IBM Corp. 2025  All Rights Reserved.
+# US Government Users Restricted Rights - Use, duplication or disclosure restricted by
+# GSA ADPSchedule Contract with IBM Corp.
+# ----------------------------------------------------------------------------------------------------
+
+from functools import partial
+from typing import Callable, Optional
+
+from wrapt import decorator
+
+from ibm_watsonx_gov.config.agentic_ai_configuration import \
+    AgenticAIConfiguration
+from ibm_watsonx_gov.entities.enums import EvaluatorFields
+from ibm_watsonx_gov.entities.metric import GenAIMetric
+from ibm_watsonx_gov.metrics.base_metric_decorator import BaseMetricDecorator
+from ibm_watsonx_gov.metrics.hap.hap_metric import HAPMetric
+
+
+class HAPDecorator(BaseMetricDecorator):
+
+    def evaluate_hap(self, func: Optional[Callable] = None,
+                     *,
+                     configuration: Optional[AgenticAIConfiguration] = None,
+                     metrics: list[GenAIMetric] = [],
+                     ) -> dict:
+        """
+        An evaluation decorator for computing hap metric on an agentic node.
+        """
+        if func is None:
+            return partial(self.evaluate_hap, configuration=configuration, metrics=metrics)
+
+        if not metrics:
+            metrics = [HAPMetric()]
+
+        @decorator
+        def wrapper(func, instance, args, kwargs):
+
+            try:
+                self.validate(func=func, metrics=metrics,
+                              valid_metric_types=(HAPMetric))
+
+                metric_inputs = [EvaluatorFields.INPUT_FIELDS]
+                original_result = self.compute_helper(func=func, args=args, kwargs=kwargs,
+                                                      configuration=configuration,
+                                                      metrics=metrics,
+                                                      metric_inputs=metric_inputs,
+                                                      metric_outputs=[])
+
+                return original_result
+            except Exception as ex:
+                raise Exception(
+                    f"There was an error while evaluating hap metric on {func.__name__},") from ex
+
+        return wrapper(func)

ibm_watsonx_gov/metrics/hap/hap_metric.py

@@ -0,0 +1,98 @@
+# ----------------------------------------------------------------------------------------------------
+# IBM Confidential
+# Licensed Materials - Property of IBM
+# 5737-H76, 5900-A3Q
+# © Copyright IBM Corp. 2025  All Rights Reserved.
+# US Government Users Restricted Rights - Use, duplication or disclosure restricted by
+# GSA ADPSchedule Contract with IBM Corp.
+# ----------------------------------------------------------------------------------------------------
+
+from typing import Annotated, Literal
+
+import pandas as pd
+from pydantic import Field
+
+from ibm_watsonx_gov.config.gen_ai_configuration import GenAIConfiguration
+from ibm_watsonx_gov.entities.enums import MetricGroup, TaskType
+from ibm_watsonx_gov.entities.evaluation_result import AggregateMetricResult
+from ibm_watsonx_gov.entities.metric import GenAIMetric
+from ibm_watsonx_gov.entities.metric_threshold import MetricThreshold
+from ibm_watsonx_gov.providers.detectors_provider import DetectorsProvider
+from ibm_watsonx_gov.utils.async_util import run_in_event_loop
+from ibm_watsonx_gov.utils.validation_util import validate_input
+
+HAP = "hap"
+
+
+class HAPMetric(GenAIMetric):
+    """
+    Defines the HAP metric class.
+
+    The HAP metric measures if there is any toxic content that contains hate, abuse, or profanity in the input.
+    It is computed using the hap model.
+
+    Examples:
+        1. Create HAP metric with default parameters and compute using metrics evaluator.
+            .. code-block:: python
+
+                metric = HAPMetric()
+                result = MetricsEvaluator().evaluate(data={"input_text": "...", metrics=[metric])
+
+        2. Create HAP metric with a custom threshold.
+            .. code-block:: python
+
+                threshold  = MetricThreshold(type="lower_limit", value=0.5)
+                metric = HAPMetric(threshold=threshold)
+    """
+    name: Annotated[Literal["hap"],
+                    Field(title="Name",
+                          description="The hap metric name.",
+                          default=HAP, frozen=True)]
+    display_name: Annotated[Literal["HAP"],
+                            Field(title="Display Name",
+                                  description="The hap metric display name.",
+                                  default="HAP", frozen=True)]
+    tasks: Annotated[list[TaskType],
+                     Field(title="Tasks",
+                           description="The list of supported tasks.",
+                           default=TaskType.values(), frozen=True)]
+    thresholds: Annotated[list[MetricThreshold],
+                          Field(title="Thresholds",
+                                description="The metric thresholds.",
+                                default=[MetricThreshold(type="upper_limit", value=0.1)])]
+    group: Annotated[MetricGroup,
+                     Field(title="Group",
+                           description="The metric group.",
+                           default=MetricGroup.CONTENT_SAFETY, frozen=True)]
+
+    async def evaluate_async(
+            self,
+            data: pd.DataFrame,
+            configuration: GenAIConfiguration,
+            **kwargs
+    ) -> list[AggregateMetricResult]:
+
+        validate_input(data.columns.to_list(), configuration)
+        provider = DetectorsProvider(configuration=configuration,
+                                     metric_name=self.name,
+                                     metric_display_name=self.display_name,
+                                     metric_method=self.method,
+                                     metric_group=self.group,
+                                     thresholds=self.thresholds,
+                                     **kwargs)
+        aggregated_metric_result = await provider.evaluate_async(data=data)
+        return aggregated_metric_result
+
+    def evaluate(
+        self,
+        data: pd.DataFrame | dict,
+        configuration: GenAIConfiguration,
+        **kwargs,
+    ):
+        # If ran in sync mode, block until it is done
+        return run_in_event_loop(
+            self.evaluate_async,
+            data=data,
+            configuration=configuration,
+            **kwargs,
+        )

ibm_watsonx_gov/metrics/hap/input_hap_metric.py

@@ -0,0 +1,104 @@
+# ----------------------------------------------------------------------------------------------------
+# IBM Confidential
+# Licensed Materials - Property of IBM
+# 5737-H76, 5900-A3Q
+# © Copyright IBM Corp. 2025  All Rights Reserved.
+# US Government Users Restricted Rights - Use, duplication or disclosure restricted by
+# GSA ADPSchedule Contract with IBM Corp.
+# ----------------------------------------------------------------------------------------------------
+
+from typing import Annotated, Literal
+
+import pandas as pd
+from pydantic import Field
+
+from ibm_watsonx_gov.config.gen_ai_configuration import GenAIConfiguration
+from ibm_watsonx_gov.entities.enums import MetricGroup, TaskType
+from ibm_watsonx_gov.entities.evaluation_result import AggregateMetricResult
+from ibm_watsonx_gov.entities.metric import GenAIMetric
+from ibm_watsonx_gov.entities.metric_threshold import MetricThreshold
+from ibm_watsonx_gov.providers.detectors_provider import DetectorsProvider
+from ibm_watsonx_gov.utils.async_util import run_in_event_loop
+from ibm_watsonx_gov.utils.validation_util import validate_input
+
+INPUT_HAP = "input_hap"
+
+
+class InputHAPMetric(GenAIMetric):
+    """
+    Defines the Input HAP metric class.
+
+    The Input HAP metric measures if there is any toxic content that contains hate, abuse, or profanity in the input.
+    It is computed using the hap model on the input data.
+
+    Examples:
+        1. Create Input HAP metric with default parameters and compute using metrics evaluator.
+            .. code-block:: python
+
+                metric = InputHAPMetric()
+                result = MetricsEvaluator().evaluate(data={"input_text": "...", metrics=[metric])
+
+        2. Create Input HAP metric with a custom threshold.
+            .. code-block:: python
+
+                threshold  = MetricThreshold(type="lower_limit", value=0.5)
+                metric = InputHAPMetric(threshold=threshold)
+    """
+    name: Annotated[Literal["input_hap"],
+                    Field(title="Name",
+                          description="The input hap metric name.",
+                          default=INPUT_HAP, frozen=True)]
+    display_name: Annotated[Literal["Input HAP"],
+                            Field(title="Display Name",
+                                  description="The input hap metric display name.",
+                                  default="Input HAP", frozen=True)]
+    tasks: Annotated[list[TaskType],
+                     Field(title="Tasks",
+                           description="The list of supported tasks.",
+                           default=TaskType.values(), frozen=True)]
+    thresholds: Annotated[list[MetricThreshold],
+                          Field(title="Thresholds",
+                                description="The metric thresholds.",
+                                default=[MetricThreshold(type="upper_limit", value=0.1)])]
+    group: Annotated[MetricGroup,
+                     Field(title="Group",
+                           description="The metric group.",
+                           default=MetricGroup.CONTENT_SAFETY, frozen=True)]
+
+    async def evaluate_async(
+            self,
+            data: pd.DataFrame,
+            configuration: GenAIConfiguration,
+            **kwargs
+    ) -> AggregateMetricResult:
+
+        validate_input(data.columns.to_list(), configuration)
+        provider = DetectorsProvider(configuration=configuration,
+                                     metric_name="hap",
+                                     metric_display_name=self.display_name,
+                                     metric_method=self.method,
+                                     metric_group=self.group,
+                                     thresholds=self.thresholds,
+                                     **kwargs)
+        aggregated_metric_result = await provider.evaluate_async(data=data)
+        # Update the metric name to input_hap
+        aggregated_metric_result.name = self.name
+        for record in aggregated_metric_result.record_level_metrics:
+            record.name = self.name
+        return aggregated_metric_result
+
+    def evaluate(
+        self,
+        data: pd.DataFrame | dict,
+        configuration: GenAIConfiguration,
+        **kwargs,
+    ):
+        # If ran in sync mode, block until it is done
+        return run_in_event_loop(
+            self.evaluate_async,
+            data=data,
+            configuration=configuration,
+            **kwargs,
+        )
+
+# Made with Bob

ibm_watsonx_gov/metrics/hap/output_hap_metric.py

@@ -0,0 +1,110 @@
+# ----------------------------------------------------------------------------------------------------
+# IBM Confidential
+# Licensed Materials - Property of IBM
+# 5737-H76, 5900-A3Q
+# © Copyright IBM Corp. 2025  All Rights Reserved.
+# US Government Users Restricted Rights - Use, duplication or disclosure restricted by
+# GSA ADPSchedule Contract with IBM Corp.
+# ----------------------------------------------------------------------------------------------------
+
+from typing import Annotated, Literal
+
+import pandas as pd
+from pydantic import Field
+
+from ibm_watsonx_gov.config.gen_ai_configuration import GenAIConfiguration
+from ibm_watsonx_gov.entities.enums import MetricGroup, TaskType
+from ibm_watsonx_gov.entities.evaluation_result import AggregateMetricResult
+from ibm_watsonx_gov.entities.metric import GenAIMetric
+from ibm_watsonx_gov.entities.metric_threshold import MetricThreshold
+from ibm_watsonx_gov.providers.detectors_provider import DetectorsProvider
+from ibm_watsonx_gov.utils.async_util import run_in_event_loop
+from ibm_watsonx_gov.utils.validation_util import validate_output
+
+OUTPUT_HAP = "output_hap"
+
+
+class OutputHAPMetric(GenAIMetric):
+    """
+    Defines the Output HAP metric class.
+
+    The Output HAP metric measures if there is any toxic content that contains hate, abuse, or profanity in the output.
+    It is computed using the hap model on the output data.
+
+    Examples:
+        1. Create Output HAP metric with default parameters and compute using metrics evaluator.
+            .. code-block:: python
+
+                metric = OutputHAPMetric()
+                result = MetricsEvaluator().evaluate(data={"generated_text": "...", metrics=[metric])
+
+        2. Create Output HAP metric with a custom threshold.
+            .. code-block:: python
+
+                threshold  = MetricThreshold(type="lower_limit", value=0.5)
+                metric = OutputHAPMetric(threshold=threshold)
+    """
+    name: Annotated[Literal["output_hap"],
+                    Field(title="Name",
+                          description="The output hap metric name.",
+                          default=OUTPUT_HAP, frozen=True)]
+    display_name: Annotated[Literal["Output HAP"],
+                            Field(title="Display Name",
+                                  description="The output hap metric display name.",
+                                  default="Output HAP", frozen=True)]
+    tasks: Annotated[list[TaskType],
+                     Field(title="Tasks",
+                           description="The list of supported tasks.",
+                           default=TaskType.values(), frozen=True)]
+    thresholds: Annotated[list[MetricThreshold],
+                          Field(title="Thresholds",
+                                description="The metric thresholds.",
+                                default=[MetricThreshold(type="upper_limit", value=0.1)])]
+    group: Annotated[MetricGroup,
+                     Field(title="Group",
+                           description="The metric group.",
+                           default=MetricGroup.CONTENT_SAFETY, frozen=True)]
+
+    async def evaluate_async(
+            self,
+            data: pd.DataFrame,
+            configuration: GenAIConfiguration,
+            **kwargs
+    ) -> AggregateMetricResult:
+
+        validate_output(data.columns.to_list(), configuration)
+
+        # Create a modified configuration that uses output_fields as input_fields
+        # This allows DetectorsProvider to process output data
+        modified_config = configuration.model_copy(deep=True)
+        modified_config.input_fields = configuration.output_fields
+
+        provider = DetectorsProvider(configuration=modified_config,
+                                     metric_name="hap",
+                                     metric_display_name=self.display_name,
+                                     metric_method=self.method,
+                                     metric_group=self.group,
+                                     thresholds=self.thresholds,
+                                     **kwargs)
+        aggregated_metric_result = await provider.evaluate_async(data=data)
+        # Update the metric name to output_hap
+        aggregated_metric_result.name = self.name
+        for record in aggregated_metric_result.record_level_metrics:
+            record.name = self.name
+        return aggregated_metric_result
+
+    def evaluate(
+        self,
+        data: pd.DataFrame | dict,
+        configuration: GenAIConfiguration,
+        **kwargs,
+    ):
+        # If ran in sync mode, block until it is done
+        return run_in_event_loop(
+            self.evaluate_async,
+            data=data,
+            configuration=configuration,
+            **kwargs,
+        )
+
+# Made with Bob

ibm_watsonx_gov/metrics/harm/__init__.py

@@ -0,0 +1,8 @@
+# ----------------------------------------------------------------------------------------------------
+# IBM Confidential
+# Licensed Materials - Property of IBM
+# 5737-H76, 5900-A3Q
+# © Copyright IBM Corp. 2025  All Rights Reserved.
+# US Government Users Restricted Rights - Use, duplication or disclosure restricted by
+# GSA ADPSchedule Contract with IBM Corp.
+# ----------------------------------------------------------------------------------------------------