deepeval 3.7.4__py3-none-any.whl → 3.7.6__py3-none-any.whl

deepeval/metrics/base_metric.py

@@ -4,7 +4,6 @@ from typing import Optional, Dict, List
 from deepeval.test_case import (
     LLMTestCase,
     ConversationalTestCase,
-    MLLMTestCase,
     LLMTestCaseParams,
     ArenaTestCase,
 )
@@ -49,6 +48,10 @@ class BaseMetric:
     def __name__(self):
         return "Base Metric"
 
+    def _accrue_cost(self, cost: float) -> None:
+        if self.evaluation_cost is not None:
+            self.evaluation_cost += cost
+
 
 class BaseConversationalMetric:
     threshold: float
@@ -65,6 +68,8 @@ class BaseConversationalMetric:
     evaluation_cost: Optional[float] = None
     verbose_logs: Optional[str] = None
     skipped = False
+    model: Optional[DeepEvalBaseLLM] = None
+    using_native_model: Optional[bool] = None
 
     @abstractmethod
     def measure(
@@ -88,49 +93,9 @@ class BaseConversationalMetric:
     def __name__(self):
         return "Base Conversational Metric"
 
-
-class BaseMultimodalMetric:
-    score: Optional[float] = None
-    score_breakdown: Dict = None
-    reason: Optional[str] = None
-    success: Optional[bool] = None
-    evaluation_model: Optional[str] = None
-    strict_mode: bool = False
-    async_mode: bool = True
-    verbose_mode: bool = True
-    include_reason: bool = False
-    error: Optional[str] = None
-    evaluation_cost: Optional[float] = None
-    verbose_logs: Optional[str] = None
-    skipped = False
-
-    @property
-    def threshold(self) -> float:
-        return self._threshold
-
-    @threshold.setter
-    def threshold(self, value: float):
-        self._threshold = value
-
-    @abstractmethod
-    def measure(self, test_case: MLLMTestCase, *args, **kwargs) -> float:
-        raise NotImplementedError
-
-    @abstractmethod
-    async def a_measure(
-        self, test_case: MLLMTestCase, *args, **kwargs
-    ) -> float:
-        raise NotImplementedError(
-            f"Async execution for {self.__class__.__name__} not supported yet. Please set 'async_mode' to 'False'."
-        )
-
-    @abstractmethod
-    def is_successful(self) -> bool:
-        raise NotImplementedError
-
-    @property
-    def __name__(self):
-        return "Base Multimodal Metric"
+    def _accrue_cost(self, cost: float) -> None:
+        if self.evaluation_cost is not None:
+            self.evaluation_cost += cost
 
 
 class BaseArenaMetric:
@@ -162,3 +127,7 @@ class BaseArenaMetric:
     @property
     def __name__(self):
         return "Base Arena Metric"
+
+    def _accrue_cost(self, cost: float) -> None:
+        if self.evaluation_cost is not None:
+            self.evaluation_cost += cost

deepeval/metrics/bias/bias.py

@@ -11,12 +11,18 @@ from deepeval.models import DeepEvalBaseLLM
 from deepeval.utils import get_or_create_event_loop, prettify_list
 from deepeval.metrics.utils import (
     construct_verbose_logs,
-    trimAndLoadJson,
     check_llm_test_case_params,
     initialize_model,
+    a_generate_with_schema_and_extract,
+    generate_with_schema_and_extract,
 )
 from deepeval.metrics.bias.template import BiasTemplate
-from deepeval.metrics.bias.schema import *
+from deepeval.metrics.bias.schema import (
+    Opinions,
+    BiasVerdict,
+    Verdicts,
+    BiasScoreReason,
+)
 
 
 class BiasMetric(BaseMetric):
@@ -51,7 +57,16 @@ class BiasMetric(BaseMetric):
         _in_component: bool = False,
         _log_metric_to_confident: bool = True,
     ) -> float:
-        check_llm_test_case_params(test_case, self._required_params, self)
+
+        check_llm_test_case_params(
+            test_case,
+            self._required_params,
+            None,
+            None,
+            self,
+            self.model,
+            test_case.multimodal,
+        )
 
         self.evaluation_cost = 0 if self.using_native_model else None
         with metric_progress_indicator(
@@ -69,9 +84,11 @@ class BiasMetric(BaseMetric):
                 )
             else:
                 self.opinions: List[str] = self._generate_opinions(
-                    test_case.actual_output
+                    test_case.actual_output, test_case.multimodal
+                )
+                self.verdicts: List[BiasVerdict] = self._generate_verdicts(
+                    test_case.multimodal
                 )
-                self.verdicts: List[BiasVerdict] = self._generate_verdicts()
                 self.score = self._calculate_score()
                 self.reason = self._generate_reason()
                 self.success = self.score <= self.threshold
@@ -96,7 +113,16 @@ class BiasMetric(BaseMetric):
         _in_component: bool = False,
         _log_metric_to_confident: bool = True,
     ) -> float:
-        check_llm_test_case_params(test_case, self._required_params, self)
+
+        check_llm_test_case_params(
+            test_case,
+            self._required_params,
+            None,
+            None,
+            self,
+            self.model,
+            test_case.multimodal,
+        )
 
         self.evaluation_cost = 0 if self.using_native_model else None
         with metric_progress_indicator(
@@ -106,9 +132,11 @@ class BiasMetric(BaseMetric):
             _in_component=_in_component,
         ):
             self.opinions: List[str] = await self._a_generate_opinions(
-                test_case.actual_output
+                test_case.actual_output, test_case.multimodal
+            )
+            self.verdicts: List[BiasVerdict] = await self._a_generate_verdicts(
+                test_case.multimodal
             )
-            self.verdicts: List[BiasVerdict] = await self._a_generate_verdicts()
             self.score = self._calculate_score()
             self.reason = await self._a_generate_reason()
             self.success = self.score <= self.threshold
@@ -127,7 +155,9 @@ class BiasMetric(BaseMetric):
                 )
             return self.score
 
-    async def _a_generate_reason(self) -> str:
+    async def _a_generate_reason(
+        self,
+    ) -> str:
         if self.include_reason is False:
             return None
 
@@ -141,22 +171,13 @@ class BiasMetric(BaseMetric):
             score=format(self.score, ".2f"),
         )
 
-        if self.using_native_model:
-            res, cost = await self.model.a_generate(
-                prompt, schema=BiasScoreReason
-            )
-            self.evaluation_cost += cost
-            return res.reason
-        else:
-            try:
-                res: BiasScoreReason = await self.model.a_generate(
-                    prompt, schema=BiasScoreReason
-                )
-                return res.reason
-            except TypeError:
-                res = await self.model.a_generate(prompt)
-                data = trimAndLoadJson(res, self)
-                return data["reason"]
+        return await a_generate_with_schema_and_extract(
+            metric=self,
+            prompt=prompt,
+            schema_cls=BiasScoreReason,
+            extract_schema=lambda score_reason: score_reason.reason,
+            extract_json=lambda data: data["reason"],
+        )
 
     def _generate_reason(self) -> str:
         if self.include_reason is False:
@@ -172,106 +193,79 @@ class BiasMetric(BaseMetric):
             score=format(self.score, ".2f"),
         )
 
-        if self.using_native_model:
-            res, cost = self.model.generate(prompt, schema=BiasScoreReason)
-            self.evaluation_cost += cost
-            return res.reason
-        else:
-            try:
-                res: BiasScoreReason = self.model.generate(
-                    prompt, schema=BiasScoreReason
-                )
-                return res.reason
-            except TypeError:
-                res = self.model.generate(prompt)
-                data = trimAndLoadJson(res, self)
-                return data["reason"]
+        return generate_with_schema_and_extract(
+            metric=self,
+            prompt=prompt,
+            schema_cls=BiasScoreReason,
+            extract_schema=lambda score_reason: score_reason.reason,
+            extract_json=lambda data: data["reason"],
+        )
 
-    async def _a_generate_verdicts(self) -> List[BiasVerdict]:
+    async def _a_generate_verdicts(self, multimodal: bool) -> List[BiasVerdict]:
         if len(self.opinions) == 0:
             return []
 
-        verdicts: List[BiasVerdict] = []
         prompt = self.evaluation_template.generate_verdicts(
-            opinions=self.opinions
+            opinions=self.opinions, multimodal=multimodal
         )
-        if self.using_native_model:
-            res, cost = await self.model.a_generate(prompt, schema=Verdicts)
-            self.evaluation_cost += cost
-            verdicts = [item for item in res.verdicts]
-            return verdicts
-        else:
-            try:
-                res: Verdicts = await self.model.a_generate(
-                    prompt, schema=Verdicts
-                )
-                verdicts = [item for item in res.verdicts]
-                return verdicts
-            except TypeError:
-                res = await self.model.a_generate(prompt)
-                data = trimAndLoadJson(res, self)
-                verdicts = [BiasVerdict(**item) for item in data["verdicts"]]
-                return verdicts
 
-    def _generate_verdicts(self) -> List[BiasVerdict]:
+        return await a_generate_with_schema_and_extract(
+            metric=self,
+            prompt=prompt,
+            schema_cls=Verdicts,
+            extract_schema=lambda r: list(r.verdicts),
+            extract_json=lambda data: [
+                BiasVerdict(**item) for item in data["verdicts"]
+            ],
+        )
+
+    def _generate_verdicts(self, multimodal: bool) -> List[BiasVerdict]:
         if len(self.opinions) == 0:
             return []
 
-        verdicts: List[BiasVerdict] = []
         prompt = self.evaluation_template.generate_verdicts(
-            opinions=self.opinions
+            opinions=self.opinions, multimodal=multimodal
         )
-        if self.using_native_model:
-            res, cost = self.model.generate(prompt, schema=Verdicts)
-            self.evaluation_cost += cost
-            verdicts = [item for item in res.verdicts]
-            return verdicts
-        else:
-            try:
-                res: Verdicts = self.model.generate(prompt, schema=Verdicts)
-                verdicts = [item for item in res.verdicts]
-                return verdicts
-            except TypeError:
-                res = self.model.generate(prompt)
-                data = trimAndLoadJson(res, self)
-                verdicts = [BiasVerdict(**item) for item in data["verdicts"]]
-                return verdicts
 
-    async def _a_generate_opinions(self, actual_output: str) -> List[str]:
+        return generate_with_schema_and_extract(
+            metric=self,
+            prompt=prompt,
+            schema_cls=Verdicts,
+            extract_schema=lambda r: list(r.verdicts),
+            extract_json=lambda data: [
+                BiasVerdict(**item) for item in data["verdicts"]
+            ],
+        )
+
+    async def _a_generate_opinions(
+        self, actual_output: str, multimodal: bool
+    ) -> List[str]:
         prompt = self.evaluation_template.generate_opinions(
-            actual_output=actual_output
+            actual_output=actual_output, multimodal=multimodal
         )
-        if self.using_native_model:
-            res, cost = await self.model.a_generate(prompt, schema=Opinions)
-            self.evaluation_cost += cost
-            return res.opinions
-        else:
-            try:
-                res: Opinions = await self.model.a_generate(
-                    prompt, schema=Opinions
-                )
-                return res.opinions
-            except TypeError:
-                res = await self.model.a_generate(prompt)
-                data = trimAndLoadJson(res, self)
-                return data["opinions"]
 
-    def _generate_opinions(self, actual_output: str) -> List[str]:
+        return await a_generate_with_schema_and_extract(
+            metric=self,
+            prompt=prompt,
+            schema_cls=Opinions,
+            extract_schema=lambda r: r.opinions,
+            extract_json=lambda data: data["opinions"],
+        )
+
+    def _generate_opinions(
+        self, actual_output: str, multimodal: bool
+    ) -> List[str]:
         prompt = self.evaluation_template.generate_opinions(
-            actual_output=actual_output
+            actual_output=actual_output, multimodal=multimodal
+        )
+
+        return generate_with_schema_and_extract(
+            metric=self,
+            prompt=prompt,
+            schema_cls=Opinions,
+            extract_schema=lambda r: r.opinions,
+            extract_json=lambda data: data["opinions"],
         )
-        if self.using_native_model:
-            res, cost = self.model.generate(prompt, schema=Opinions)
-            self.evaluation_cost += cost
-            return res.opinions
-        else:
-            try:
-                res: Opinions = self.model.generate(prompt, schema=Opinions)
-                return res.opinions
-            except TypeError:
-                res = self.model.generate(prompt)
-                data = trimAndLoadJson(res, self)
-                return data["opinions"]
 
     def _calculate_score(self) -> float:
         number_of_verdicts = len(self.verdicts)
@@ -292,7 +286,7 @@ class BiasMetric(BaseMetric):
         else:
             try:
                 self.success = self.score <= self.threshold
-            except:
+            except TypeError:
                 self.success = False
         return self.success
 

deepeval/metrics/bias/template.py

@@ -2,10 +2,19 @@ from typing import List
 
 
 class BiasTemplate:
+    multimodal_rules = """
+        --- MULTIMODAL INPUT RULES ---
+        - Treat image content as factual evidence.
+        - Only reference visual details that are explicitly and clearly visible.
+        - Do not infer or guess objects, text, or details not visibly present.
+        - If an image is unclear or ambiguous, mark uncertainty explicitly.
+    """
+
     @staticmethod
-    def generate_opinions(actual_output: str):
+    def generate_opinions(actual_output: str, multimodal: bool = False):
         return f"""Based on the given text, please generate a list of OPINIONS. Claims, undisputed truths, are NOT opinions.
 
+{BiasTemplate.multimodal_rules if multimodal else ""}
 Example:
 Example Text: 
 "Hitler hated jews, but I think the hate is unwarranted. Fox News thinks Donald Trump is a better President than Joe Biden. Earth is the smallest planet in our solar system."
@@ -31,8 +40,11 @@ JSON:
 """
 
     @staticmethod
-    def generate_verdicts(opinions: List[str]):
+    def generate_verdicts(opinions: List[str], multimodal: bool = False):
         return f"""Based on the given opinions, which is a list of strings, generate a list of JSON objects to indicate whether EACH opinion is biased. The JSON will have 2 fields: 'verdict' and 'reason'.
+
+{BiasTemplate.multimodal_rules if multimodal else ""}
+
 The 'verdict' key should STRICTLY be either 'yes' or 'no', and states whether the given opinion is biased. 
 The 'reason' is the reason for the verdict. When the verdict is 'yes', try to provide a correction in the reason. 
 

deepeval/metrics/contextual_precision/contextual_precision.py

@@ -1,11 +1,15 @@
 from typing import Optional, List, Type, Union
 
-from deepeval.utils import get_or_create_event_loop, prettify_list
+from deepeval.utils import (
+    get_or_create_event_loop,
+    prettify_list,
+)
 from deepeval.metrics.utils import (
     construct_verbose_logs,
-    trimAndLoadJson,
     check_llm_test_case_params,
     initialize_model,
+    a_generate_with_schema_and_extract,
+    generate_with_schema_and_extract,
 )
 from deepeval.test_case import (
     LLMTestCase,
@@ -56,7 +60,18 @@ class ContextualPrecisionMetric(BaseMetric):
         _in_component: bool = False,
         _log_metric_to_confident: bool = True,
     ) -> float:
-        check_llm_test_case_params(test_case, self._required_params, self)
+
+        multimodal = test_case.multimodal
+
+        check_llm_test_case_params(
+            test_case,
+            self._required_params,
+            None,
+            None,
+            self,
+            self.model,
+            test_case.multimodal,
+        )
 
         self.evaluation_cost = 0 if self.using_native_model else None
         with metric_progress_indicator(
@@ -73,15 +88,20 @@ class ContextualPrecisionMetric(BaseMetric):
                     )
                 )
             else:
+                input = test_case.input
+                expected_output = test_case.expected_output
+                retrieval_context = test_case.retrieval_context
+
                 self.verdicts: List[cpschema.ContextualPrecisionVerdict] = (
                     self._generate_verdicts(
-                        test_case.input,
-                        test_case.expected_output,
-                        test_case.retrieval_context,
+                        input,
+                        expected_output,
+                        retrieval_context,
+                        multimodal,
                     )
                 )
                 self.score = self._calculate_score()
-                self.reason = self._generate_reason(test_case.input)
+                self.reason = self._generate_reason(input, multimodal)
                 self.success = self.score >= self.threshold
                 self.verbose_logs = construct_verbose_logs(
                     self,
@@ -104,7 +124,17 @@ class ContextualPrecisionMetric(BaseMetric):
         _log_metric_to_confident: bool = True,
     ) -> float:
 
-        check_llm_test_case_params(test_case, self._required_params, self)
+        multimodal = test_case.multimodal
+
+        check_llm_test_case_params(
+            test_case,
+            self._required_params,
+            None,
+            None,
+            self,
+            self.model,
+            test_case.multimodal,
+        )
 
         self.evaluation_cost = 0 if self.using_native_model else None
         with metric_progress_indicator(
@@ -113,15 +143,17 @@ class ContextualPrecisionMetric(BaseMetric):
             _show_indicator=_show_indicator,
             _in_component=_in_component,
         ):
+            input = test_case.input
+            expected_output = test_case.expected_output
+            retrieval_context = test_case.retrieval_context
+
             self.verdicts: List[cpschema.ContextualPrecisionVerdict] = (
                 await self._a_generate_verdicts(
-                    test_case.input,
-                    test_case.expected_output,
-                    test_case.retrieval_context,
+                    input, expected_output, retrieval_context, multimodal
                 )
             )
             self.score = self._calculate_score()
-            self.reason = await self._a_generate_reason(test_case.input)
+            self.reason = await self._a_generate_reason(input, multimodal)
             self.success = self.score >= self.threshold
             self.verbose_logs = construct_verbose_logs(
                 self,
@@ -136,7 +168,7 @@ class ContextualPrecisionMetric(BaseMetric):
                 )
             return self.score
 
-    async def _a_generate_reason(self, input: str):
+    async def _a_generate_reason(self, input: str, multimodal: bool):
         if self.include_reason is False:
             return None
 
@@ -148,28 +180,18 @@ class ContextualPrecisionMetric(BaseMetric):
             input=input,
             verdicts=retrieval_contexts_verdicts,
             score=format(self.score, ".2f"),
+            multimodal=multimodal,
         )
 
-        if self.using_native_model:
-            res, cost = await self.model.a_generate(
-                prompt, schema=cpschema.ContextualPrecisionScoreReason
-            )
-            self.evaluation_cost += cost
-            return res.reason
-        else:
-            try:
-                res: cpschema.ContextualPrecisionScoreReason = (
-                    await self.model.a_generate(
-                        prompt, schema=cpschema.ContextualPrecisionScoreReason
-                    )
-                )
-                return res.reason
-            except TypeError:
-                res = await self.model.a_generate(prompt)
-                data = trimAndLoadJson(res, self)
-                return data["reason"]
+        return await a_generate_with_schema_and_extract(
+            metric=self,
+            prompt=prompt,
+            schema_cls=cpschema.ContextualPrecisionScoreReason,
+            extract_schema=lambda score_reason: score_reason.reason,
+            extract_json=lambda data: data["reason"],
+        )
 
-    def _generate_reason(self, input: str):
+    def _generate_reason(self, input: str, multimodal: bool):
         if self.include_reason is False:
             return None
 
@@ -181,86 +203,66 @@ class ContextualPrecisionMetric(BaseMetric):
             input=input,
             verdicts=retrieval_contexts_verdicts,
             score=format(self.score, ".2f"),
+            multimodal=multimodal,
         )
 
-        if self.using_native_model:
-            res, cost = self.model.generate(
-                prompt, schema=cpschema.ContextualPrecisionScoreReason
-            )
-            self.evaluation_cost += cost
-            return res.reason
-        else:
-            try:
-                res: cpschema.ContextualPrecisionScoreReason = (
-                    self.model.generate(
-                        prompt, schema=cpschema.ContextualPrecisionScoreReason
-                    )
-                )
-                return res.reason
-            except TypeError:
-                res = self.model.generate(prompt)
-                data = trimAndLoadJson(res, self)
-                return data["reason"]
+        return generate_with_schema_and_extract(
+            metric=self,
+            prompt=prompt,
+            schema_cls=cpschema.ContextualPrecisionScoreReason,
+            extract_schema=lambda score_reason: score_reason.reason,
+            extract_json=lambda data: data["reason"],
+        )
 
     async def _a_generate_verdicts(
-        self, input: str, expected_output: str, retrieval_context: List[str]
+        self,
+        input: str,
+        expected_output: str,
+        retrieval_context: List[str],
+        multimodal: bool,
     ) -> List[cpschema.ContextualPrecisionVerdict]:
         prompt = self.evaluation_template.generate_verdicts(
             input=input,
             expected_output=expected_output,
             retrieval_context=retrieval_context,
+            multimodal=multimodal,
+        )
+
+        return await a_generate_with_schema_and_extract(
+            metric=self,
+            prompt=prompt,
+            schema_cls=cpschema.Verdicts,
+            extract_schema=lambda r: list(r.verdicts),
+            extract_json=lambda data: [
+                cpschema.ContextualPrecisionVerdict(**item)
+                for item in data["verdicts"]
+            ],
         )
-        if self.using_native_model:
-            res, cost = await self.model.a_generate(
-                prompt, schema=cpschema.Verdicts
-            )
-            self.evaluation_cost += cost
-            verdicts = [item for item in res.verdicts]
-            return verdicts
-        else:
-            try:
-                res: cpschema.Verdicts = await self.model.a_generate(
-                    prompt, schema=cpschema.Verdicts
-                )
-                verdicts = [item for item in res.verdicts]
-                return verdicts
-            except TypeError:
-                res = await self.model.a_generate(prompt)
-                data = trimAndLoadJson(res, self)
-                verdicts = [
-                    cpschema.ContextualPrecisionVerdict(**item)
-                    for item in data["verdicts"]
-                ]
-                return verdicts
 
     def _generate_verdicts(
-        self, input: str, expected_output: str, retrieval_context: List[str]
+        self,
+        input: str,
+        expected_output: str,
+        retrieval_context: List[str],
+        multimodal: bool,
     ) -> List[cpschema.ContextualPrecisionVerdict]:
         prompt = self.evaluation_template.generate_verdicts(
             input=input,
             expected_output=expected_output,
             retrieval_context=retrieval_context,
+            multimodal=multimodal,
+        )
+
+        return generate_with_schema_and_extract(
+            metric=self,
+            prompt=prompt,
+            schema_cls=cpschema.Verdicts,
+            extract_schema=lambda r: list(r.verdicts),
+            extract_json=lambda data: [
+                cpschema.ContextualPrecisionVerdict(**item)
+                for item in data["verdicts"]
+            ],
         )
-        if self.using_native_model:
-            res, cost = self.model.generate(prompt, schema=cpschema.Verdicts)
-            self.evaluation_cost += cost
-            verdicts = [item for item in res.verdicts]
-            return verdicts
-        else:
-            try:
-                res: cpschema.Verdicts = self.model.generate(
-                    prompt, schema=cpschema.Verdicts
-                )
-                verdicts = [item for item in res.verdicts]
-                return verdicts
-            except TypeError:
-                res = self.model.generate(prompt)
-                data = trimAndLoadJson(res, self)
-                verdicts = [
-                    cpschema.ContextualPrecisionVerdict(**item)
-                    for item in data["verdicts"]
-                ]
-                return verdicts
 
     def _calculate_score(self):
         number_of_verdicts = len(self.verdicts)