deepeval 3.7.5__py3-none-any.whl → 3.7.7__py3-none-any.whl

deepeval/metrics/argument_correctness/argument_correctness.py

@@ -3,9 +3,10 @@ from typing import Optional, List, Type, Union
 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,
@@ -18,7 +19,11 @@ from deepeval.metrics.argument_correctness.template import (
     ArgumentCorrectnessTemplate,
 )
 from deepeval.metrics.indicator import metric_progress_indicator
-from deepeval.metrics.argument_correctness.schema import *
+from deepeval.metrics.argument_correctness.schema import (
+    ArgumentCorrectnessVerdict,
+    Verdicts,
+    ArgumentCorrectnessScoreReason,
+)
 from deepeval.metrics.api import metric_data_manager
 
 
@@ -57,7 +62,15 @@ class ArgumentCorrectnessMetric(BaseMetric):
         _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(
@@ -81,11 +94,15 @@ class ArgumentCorrectnessMetric(BaseMetric):
                 else:
                     self.verdicts: List[ArgumentCorrectnessVerdict] = (
                         self._generate_verdicts(
-                            test_case.input, test_case.tools_called
+                            test_case.input,
+                            test_case.tools_called,
+                            test_case.multimodal,
                         )
                     )
                     self.score = self._calculate_score()
-                    self.reason = self._generate_reason(test_case.input)
+                    self.reason = self._generate_reason(
+                        test_case.input, test_case.multimodal
+                    )
                 self.success = self.score >= self.threshold
                 self.verbose_logs = construct_verbose_logs(
                     self,
@@ -108,7 +125,15 @@ class ArgumentCorrectnessMetric(BaseMetric):
         _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(
@@ -124,11 +149,15 @@ class ArgumentCorrectnessMetric(BaseMetric):
             else:
                 self.verdicts: List[ArgumentCorrectnessVerdict] = (
                     await self._a_generate_verdicts(
-                        test_case.input, test_case.tools_called
+                        test_case.input,
+                        test_case.tools_called,
+                        test_case.multimodal,
                     )
                 )
                 self.score = self._calculate_score()
-                self.reason = await self._a_generate_reason(test_case.input)
+                self.reason = await self._a_generate_reason(
+                    test_case.input, test_case.multimodal
+                )
             self.success = self.score >= self.threshold
             self.verbose_logs = construct_verbose_logs(
                 self,
@@ -143,7 +172,7 @@ class ArgumentCorrectnessMetric(BaseMetric):
                 )
             return self.score
 
-    async def _a_generate_reason(self, input: str) -> str:
+    async def _a_generate_reason(self, input: str, multimodal: bool) -> str:
         if self.include_reason is False:
             return None
 
@@ -156,27 +185,18 @@ class ArgumentCorrectnessMetric(BaseMetric):
             incorrect_tool_calls_reasons=incorrect_tool_calls_reasons,
             input=input,
             score=format(self.score, ".2f"),
+            multimodal=multimodal,
+        )
+
+        return await a_generate_with_schema_and_extract(
+            metric=self,
+            prompt=prompt,
+            schema_cls=ArgumentCorrectnessScoreReason,
+            extract_schema=lambda score_reason: score_reason.reason,
+            extract_json=lambda data: data["reason"],
         )
-        if self.using_native_model:
-            res, cost = await self.model.a_generate(
-                prompt, schema=ArgumentCorrectnessScoreReason
-            )
-            self.evaluation_cost += cost
-            return res.reason
-        else:
-            try:
-                res: ArgumentCorrectnessScoreReason = (
-                    await self.model.a_generate(
-                        prompt=prompt, schema=ArgumentCorrectnessScoreReason
-                    )
-                )
-                return res.reason
-            except TypeError:
-                res = await self.model.a_generate(prompt)
-                data = trimAndLoadJson(res, self)
-                return data["reason"]
 
-    def _generate_reason(self, input: str) -> str:
+    def _generate_reason(self, input: str, multimodal: bool) -> str:
         if self.include_reason is False:
             return None
 
@@ -189,76 +209,50 @@ class ArgumentCorrectnessMetric(BaseMetric):
             incorrect_tool_calls_reasons=incorrect_tool_calls_reasons,
             input=input,
             score=format(self.score, ".2f"),
+            multimodal=multimodal,
         )
 
-        if self.using_native_model:
-            res, cost = self.model.generate(
-                prompt, schema=ArgumentCorrectnessScoreReason
-            )
-            self.evaluation_cost += cost
-            return res.reason
-        else:
-            try:
-                res: ArgumentCorrectnessScoreReason = self.model.generate(
-                    prompt=prompt, schema=ArgumentCorrectnessScoreReason
-                )
-                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=ArgumentCorrectnessScoreReason,
+            extract_schema=lambda score_reason: score_reason.reason,
+            extract_json=lambda data: data["reason"],
+        )
 
     async def _a_generate_verdicts(
-        self,
-        input: str,
-        tools_called: List[ToolCall],
+        self, input: str, tools_called: List[ToolCall], multimodal: bool
     ) -> List[ArgumentCorrectnessVerdict]:
         prompt = self.evaluation_template.generate_verdicts(
-            input=input,
-            tools_called=tools_called,
+            input=input, tools_called=tools_called, multimodal=multimodal
+        )
+
+        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: [
+                ArgumentCorrectnessVerdict(**item) for item in data["verdicts"]
+            ],
         )
-        if self.using_native_model:
-            res, cost = await self.model.a_generate(prompt, schema=Verdicts)
-            self.evaluation_cost += cost
-            return [item for item in res.verdicts]
-        else:
-            try:
-                res: Verdicts = await self.model.a_generate(
-                    prompt, schema=Verdicts
-                )
-                return [item for item in res.verdicts]
-            except TypeError:
-                res = await self.model.a_generate(prompt)
-                data = trimAndLoadJson(res, self)
-                return [
-                    ArgumentCorrectnessVerdict(**item)
-                    for item in data["verdicts"]
-                ]
 
     def _generate_verdicts(
-        self,
-        input: str,
-        tools_called: List[ToolCall],
+        self, input: str, tools_called: List[ToolCall], multimodal: bool
     ) -> List[ArgumentCorrectnessVerdict]:
         prompt = self.evaluation_template.generate_verdicts(
-            input=input,
-            tools_called=tools_called,
+            input=input, tools_called=tools_called, multimodal=multimodal
+        )
+
+        return generate_with_schema_and_extract(
+            metric=self,
+            prompt=prompt,
+            schema_cls=Verdicts,
+            extract_schema=lambda r: list(r.verdicts),
+            extract_json=lambda data: [
+                ArgumentCorrectnessVerdict(**item) for item in data["verdicts"]
+            ],
         )
-        if self.using_native_model:
-            res, cost = self.model.generate(prompt, schema=Verdicts)
-            self.evaluation_cost += cost
-            return [item for item in res.verdicts]
-        else:
-            try:
-                res: Verdicts = self.model.generate(prompt, schema=Verdicts)
-                return [item for item in res.verdicts]
-            except TypeError:
-                res = self.model.generate(prompt)
-                data = trimAndLoadJson(res, self)
-                return [
-                    ArgumentCorrectnessVerdict(**item)
-                    for item in data["verdicts"]
-                ]
 
     def _calculate_score(self):
         number_of_verdicts = len(self.verdicts)
@@ -279,7 +273,7 @@ class ArgumentCorrectnessMetric(BaseMetric):
         else:
             try:
                 self.success = self.score >= self.threshold
-            except:
+            except TypeError:
                 self.success = False
         return self.success
 

deepeval/metrics/argument_correctness/template.py

@@ -4,8 +4,18 @@ import textwrap
 
 
 class ArgumentCorrectnessTemplate:
+    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_verdicts(input: str, tools_called: List[ToolCall]):
+    def generate_verdicts(
+        input: str, tools_called: List[ToolCall], multimodal: bool = False
+    ):
 
         stringified_tools_called = repr(tools_called)
 
@@ -19,6 +29,8 @@ class ArgumentCorrectnessTemplate:
             Provide a 'reason' ONLY if the answer is 'no'. 
             If there is no input parameter, answer 'no' for the verdict and provide the reason as "No input parameter provided".
 
+            {ArgumentCorrectnessTemplate.multimodal_rules if multimodal else ""}
+
             **
             IMPORTANT: Please make sure to only return in valid and parseable JSON format, with the 'verdicts' key mapping to a list of JSON objects. Ensure all strings are closed appropriately. Repair any invalid JSON before you output it.
             Example input: 
@@ -92,12 +104,17 @@ class ArgumentCorrectnessTemplate:
 
     @staticmethod
     def generate_reason(
-        incorrect_tool_calls_reasons: List[str], input: str, score: float
+        incorrect_tool_calls_reasons: List[str],
+        input: str,
+        score: float,
+        multimodal: bool = False,
     ):
         return textwrap.dedent(
             f"""Given the argument correctness score, the list of reasons of incorrect tool calls, and the input, provide a CONCISE reason for the score. Explain why it is not higher, but also why it is at its current score. You can mention tool calls or input, but do not mention an output or a response.
             If there is nothing incorrect, just say something positive with an upbeat encouraging tone (but don't overdo it otherwise it gets annoying).
 
+            {ArgumentCorrectnessTemplate.multimodal_rules if multimodal else ""}
+
             **
             IMPORTANT: Please make sure to only return in JSON format, with the 'reason' key providing the reason. Ensure all strings are closed appropriately. Repair any invalid JSON before you output it.
 

deepeval/metrics/base_metric.py

@@ -48,6 +48,12 @@ class BaseMetric:
     def __name__(self):
         return "Base Metric"
 
+    def _accrue_cost(self, cost: float) -> None:
+        if self.evaluation_cost is not None and cost is not None:
+            self.evaluation_cost += cost
+        else:
+            self.evaluation_cost = None
+
 
 class BaseConversationalMetric:
     threshold: float
@@ -64,6 +70,8 @@ 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(
@@ -87,47 +95,11 @@ 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: LLMTestCase, *args, **kwargs) -> float:
-        raise NotImplementedError
-
-    @abstractmethod
-    async def a_measure(self, test_case: LLMTestCase, *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 and cost is not None:
+            self.evaluation_cost += cost
+        else:
+            self.evaluation_cost = None
 
 
 class BaseArenaMetric:
@@ -159,3 +131,9 @@ class BaseArenaMetric:
     @property
     def __name__(self):
         return "Base Arena Metric"
+
+    def _accrue_cost(self, cost: float) -> None:
+        if self.evaluation_cost is not None and cost is not None:
+            self.evaluation_cost += cost
+        else:
+            self.evaluation_cost = None

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