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

deepeval/metrics/conversational_dag/nodes.py

@@ -8,7 +8,11 @@ from deepeval.metrics.conversational_g_eval.conversational_g_eval import (
     ConversationalGEval,
 )
 from deepeval.metrics.g_eval.utils import CONVERSATIONAL_G_EVAL_PARAMS
-from deepeval.metrics.utils import copy_metrics, trimAndLoadJson
+from deepeval.metrics.utils import (
+    copy_metrics,
+    a_generate_with_schema_and_extract,
+    generate_with_schema_and_extract,
+)
 from deepeval.test_case import (
     ConversationalTestCase,
     TurnParams,
@@ -263,20 +267,14 @@ class ConversationalVerdictNode(ConversationalBaseNode):
             score=metric.score,
             name=metric.__name__,
         )
-        if metric.using_native_model:
-            res, cost = metric.model.generate(prompt, schema=MetricScoreReason)
-            metric.evaluation_cost += cost
-        else:
-            try:
-                res: MetricScoreReason = metric.model.generate(
-                    prompt, schema=MetricScoreReason
-                )
-            except TypeError:
-                res = metric.model.generate(prompt)
-                data = trimAndLoadJson(res, self)
-                res = MetricScoreReason(**data)
 
-        return res.reason
+        return generate_with_schema_and_extract(
+            metric=metric,
+            prompt=prompt,
+            schema_cls=MetricScoreReason,
+            extract_schema=lambda score_reason: score_reason.reason,
+            extract_json=lambda data: data["reason"],
+        )
 
     async def _a_generate_reason(self, metric: BaseConversationalMetric):
         prompt = ConversationalVerdictNodeTemplate.generate_reason(
@@ -284,22 +282,14 @@ class ConversationalVerdictNode(ConversationalBaseNode):
             score=metric.score,
             name=metric.__name__,
         )
-        if metric.using_native_model:
-            res, cost = await metric.model.a_generate(
-                prompt, schema=MetricScoreReason
-            )
-            metric.evaluation_cost += cost
-        else:
-            try:
-                res: MetricScoreReason = await metric.model.a_generate(
-                    prompt, schema=MetricScoreReason
-                )
-            except TypeError:
-                res = await metric.model.a_generate(prompt)
-                data = trimAndLoadJson(res, self)
-                res = MetricScoreReason(**data)
 
-        return res.reason
+        return await a_generate_with_schema_and_extract(
+            metric=metric,
+            prompt=prompt,
+            schema_cls=MetricScoreReason,
+            extract_schema=lambda score_reason: score_reason.reason,
+            extract_json=lambda data: data["reason"],
+        )
 
 
 @dataclass
@@ -372,20 +362,14 @@ class ConversationalTaskNode(ConversationalBaseNode):
             instructions=self.instructions,
             text=text,
         )
-        if metric.using_native_model:
-            res, cost = metric.model.generate(prompt, schema=TaskNodeOutput)
-            metric.evaluation_cost += cost
-            self._output = res.output
-        else:
-            try:
-                res: TaskNodeOutput = metric.model.generate(
-                    prompt, schema=TaskNodeOutput
-                )
-                self._output = res.output
-            except TypeError:
-                res = metric.model.generate(prompt)
-                data = trimAndLoadJson(res, self)
-                self._output = TaskNodeOutput(**data).output
+
+        self._output = generate_with_schema_and_extract(
+            metric=metric,
+            prompt=prompt,
+            schema_cls=TaskNodeOutput,
+            extract_schema=lambda s: s.output,
+            extract_json=lambda data: data["output"],
+        )
 
         metric._verbose_steps.append(
             construct_node_verbose_log(self, self._depth)
@@ -439,23 +423,14 @@ class ConversationalTaskNode(ConversationalBaseNode):
             instructions=self.instructions,
             text=text,
         )
-        if metric.using_native_model:
-            res, cost = await metric.model.a_generate(
-                prompt, schema=TaskNodeOutput
-            )
-            metric.evaluation_cost += cost
-            self._output = res.output
-        else:
-            try:
-                res: TaskNodeOutput = await metric.model.a_generate(
-                    prompt, schema=TaskNodeOutput
-                )
-                self._output = res.output
-            except TypeError:
-                res = await metric.model.a_generate(prompt)
-                data = trimAndLoadJson(res, self)
-                self._output = TaskNodeOutput(**data).output
 
+        self._output = await a_generate_with_schema_and_extract(
+            metric=metric,
+            prompt=prompt,
+            schema_cls=TaskNodeOutput,
+            extract_schema=lambda s: s.output,
+            extract_json=lambda data: data["output"],
+        )
         metric._verbose_steps.append(
             construct_node_verbose_log(self, self._depth)
         )
@@ -559,22 +534,14 @@ class ConversationalBinaryJudgementNode(ConversationalBaseNode):
             criteria=self.criteria,
             text=text,
         )
-        if metric.using_native_model:
-            res, cost = metric.model.generate(
-                prompt, schema=BinaryJudgementVerdict
-            )
-            metric.evaluation_cost += cost
-            self._verdict = res
-        else:
-            try:
-                res: BinaryJudgementVerdict = metric.model.generate(
-                    prompt, schema=BinaryJudgementVerdict
-                )
-                self._verdict = res
-            except TypeError:
-                res = metric.model.generate(prompt)
-                data = trimAndLoadJson(res, self)
-                self._verdict = BinaryJudgementVerdict(**data)
+
+        self._verdict = generate_with_schema_and_extract(
+            metric=metric,
+            prompt=prompt,
+            schema_cls=BinaryJudgementVerdict,
+            extract_schema=lambda s: s,
+            extract_json=lambda data: BinaryJudgementVerdict(**data),
+        )
 
         metric._verbose_steps.append(
             construct_node_verbose_log(self, self._depth)
@@ -623,22 +590,14 @@ class ConversationalBinaryJudgementNode(ConversationalBaseNode):
             criteria=self.criteria,
             text=text,
         )
-        if metric.using_native_model:
-            res, cost = await metric.model.a_generate(
-                prompt, schema=BinaryJudgementVerdict
-            )
-            metric.evaluation_cost += cost
-            self._verdict = res
-        else:
-            try:
-                res: BinaryJudgementVerdict = await metric.model.a_generate(
-                    prompt, schema=BinaryJudgementVerdict
-                )
-                self._verdict = res
-            except TypeError:
-                res = await metric.model.a_generate(prompt)
-                data = trimAndLoadJson(res, self)
-                self._verdict = BinaryJudgementVerdict(**data)
+
+        self._verdict = await a_generate_with_schema_and_extract(
+            metric=metric,
+            prompt=prompt,
+            schema_cls=BinaryJudgementVerdict,
+            extract_schema=lambda s: s,
+            extract_json=lambda data: BinaryJudgementVerdict(**data),
+        )
 
         metric._verbose_steps.append(
             construct_node_verbose_log(self, self._depth)
@@ -753,22 +712,14 @@ class ConversationalNonBinaryJudgementNode(ConversationalBaseNode):
         prompt = ConversationalNonBinaryJudgementTemplate.generate_non_binary_verdict(
             criteria=self.criteria, text=text, options=self._verdict_options
         )
-        if metric.using_native_model:
-            res, cost = metric.model.generate(
-                prompt, schema=self._verdict_schema
-            )
-            metric.evaluation_cost += cost
-            self._verdict = res
-        else:
-            try:
-                res: self._verdict_schema = metric.model.generate(
-                    prompt, schema=self._verdict_schema
-                )
-                self._verdict = res
-            except TypeError:
-                res = metric.model.generate(prompt)
-                data = trimAndLoadJson(res, self)
-                self._verdict = self._verdict_schema(**data)
+
+        self._verdict = generate_with_schema_and_extract(
+            metric=metric,
+            prompt=prompt,
+            schema_cls=self._verdict_schema,
+            extract_schema=lambda s: s,
+            extract_json=lambda data: self._verdict_schema(**data),
+        )
 
         metric._verbose_steps.append(
             construct_node_verbose_log(self, self._depth)
@@ -816,22 +767,14 @@ class ConversationalNonBinaryJudgementNode(ConversationalBaseNode):
         prompt = ConversationalNonBinaryJudgementTemplate.generate_non_binary_verdict(
             criteria=self.criteria, text=text, options=self._verdict_options
         )
-        if metric.using_native_model:
-            res, cost = await metric.model.a_generate(
-                prompt, schema=self._verdict_schema
-            )
-            metric.evaluation_cost += cost
-            self._verdict = res
-        else:
-            try:
-                res: self._verdict_schema = await metric.model.a_generate(
-                    prompt, schema=self._verdict_schema
-                )
-                self._verdict = res
-            except TypeError:
-                res = await metric.model.a_generate(prompt)
-                data = trimAndLoadJson(res, self)
-                self._verdict = self._verdict_schema(**data)
+
+        self._verdict = await a_generate_with_schema_and_extract(
+            metric=metric,
+            prompt=prompt,
+            schema_cls=self._verdict_schema,
+            extract_schema=lambda s: s,
+            extract_json=lambda data: self._verdict_schema(**data),
+        )
 
         metric._verbose_steps.append(
             construct_node_verbose_log(self, self._depth)

deepeval/metrics/conversational_dag/templates.py

@@ -2,7 +2,17 @@ from typing import List
 from textwrap import dedent
 
 
+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.
+"""
+
+
 class ConversationalVerdictNodeTemplate:
+
     @staticmethod
     def generate_reason(verbose_steps: List[str], score: float, name: str):
         return dedent(
@@ -40,6 +50,8 @@ class ConversationalTaskNodeTemplate:
         return dedent(
             f"""You are given a set of task instructions and a full conversation between a user and an assistant.
 
+                {multimodal_rules}
+
                 Instructions:
                 {instructions}
 
@@ -67,6 +79,8 @@ class ConversationalBinaryJudgementTemplate:
 
                 Below is the full conversation you should evaluate. Consider dialogue context, speaker roles, and how responses were handled.
 
+                {multimodal_rules}
+
                 Full Conversation:
                 {text}
 
@@ -96,6 +110,8 @@ class ConversationalNonBinaryJudgementTemplate:
 
                 You are evaluating the following conversation. Choose one of the options that best reflects the assistant's behavior.
 
+                {multimodal_rules}
+
                 Options: {options}
 
                 Full Conversation:

deepeval/metrics/conversational_g_eval/conversational_g_eval.py

@@ -26,6 +26,8 @@ from deepeval.metrics.utils import (
     trimAndLoadJson,
     initialize_model,
     convert_turn_to_dict,
+    a_generate_with_schema_and_extract,
+    generate_with_schema_and_extract,
 )
 from deepeval.models import DeepEvalBaseLLM
 from deepeval.metrics.indicator import metric_progress_indicator
@@ -89,8 +91,14 @@ class ConversationalGEval(BaseConversationalMetric):
         _in_component: bool = False,
         _log_metric_to_confident: bool = True,
     ) -> float:
+        multimodal = test_case.multimodal
         check_conversational_test_case_params(
-            test_case, self.evaluation_params, self
+            test_case,
+            self.evaluation_params,
+            self,
+            False,
+            self.model,
+            multimodal,
         )
 
         self.evaluation_cost = 0 if self.using_native_model else None
@@ -143,8 +151,14 @@ class ConversationalGEval(BaseConversationalMetric):
         _in_component: bool = False,
         _log_metric_to_confident: bool = True,
     ) -> float:
+        multimodal = test_case.multimodal
         check_conversational_test_case_params(
-            test_case, self.evaluation_params, self
+            test_case,
+            self.evaluation_params,
+            self,
+            False,
+            self.model,
+            multimodal,
         )
 
         self.evaluation_cost = 0 if self.using_native_model else None
@@ -192,22 +206,13 @@ class ConversationalGEval(BaseConversationalMetric):
         prompt = self.evaluation_template.generate_evaluation_steps(
             criteria=self.criteria, parameters=g_eval_params_str
         )
-        if self.using_native_model:
-            res, cost = await self.model.a_generate(
-                prompt, schema=cgschema.Steps
-            )
-            self.evaluation_cost += cost
-            return res.steps
-        else:
-            try:
-                res: cgschema.Steps = await self.model.a_generate(
-                    prompt, schema=cgschema.Steps
-                )
-                return res.steps
-            except TypeError:
-                res = await self.model.a_generate(prompt)
-                data = trimAndLoadJson(res, self)
-                return data["steps"]
+        return await a_generate_with_schema_and_extract(
+            metric=self,
+            prompt=prompt,
+            schema_cls=cgschema.Steps,
+            extract_schema=lambda s: s.steps,
+            extract_json=lambda data: data["steps"],
+        )
 
     def _generate_evaluation_steps(self) -> List[str]:
         if self.evaluation_steps:
@@ -219,20 +224,13 @@ class ConversationalGEval(BaseConversationalMetric):
         prompt = self.evaluation_template.generate_evaluation_steps(
             criteria=self.criteria, parameters=g_eval_params_str
         )
-        if self.using_native_model:
-            res, cost = self.model.generate(prompt, schema=cgschema.Steps)
-            self.evaluation_cost += cost
-            return res.steps
-        else:
-            try:
-                res: cgschema.Steps = self.model.generate(
-                    prompt, schema=cgschema.Steps
-                )
-                return res.steps
-            except TypeError:
-                res = self.model.generate(prompt)
-                data = trimAndLoadJson(res, self)
-                return data["steps"]
+        return generate_with_schema_and_extract(
+            metric=self,
+            prompt=prompt,
+            schema_cls=cgschema.Steps,
+            extract_schema=lambda s: s.steps,
+            extract_json=lambda data: data["steps"],
+        )
 
     async def _a_evaluate(
         self, test_case: ConversationalTestCase
@@ -269,7 +267,8 @@ class ConversationalGEval(BaseConversationalMetric):
             res, cost = await self.model.a_generate_raw_response(
                 prompt, top_logprobs=20
             )
-            self.evaluation_cost += cost
+
+            self._accrue_cost(cost)
             data = trimAndLoadJson(res.choices[0].message.content, self)
 
             reason = data["reason"]
@@ -287,22 +286,13 @@ class ConversationalGEval(BaseConversationalMetric):
         except (
             AttributeError
         ):  # This catches the case where a_generate_raw_response doesn't exist.
-            if self.using_native_model:
-                res, cost = await self.model.a_generate(
-                    prompt, schema=cgschema.ReasonScore
-                )
-                self.evaluation_cost += cost
-                return res.score, res.reason
-            else:
-                try:
-                    res: cgschema.ReasonScore = await self.model.a_generate(
-                        prompt, schema=cgschema.ReasonScore
-                    )
-                    return res.score, res.reason
-                except TypeError:
-                    res = await self.model.a_generate(prompt)
-                    data = trimAndLoadJson(res, self)
-                    return data["score"], data["reason"]
+            return await a_generate_with_schema_and_extract(
+                metric=self,
+                prompt=prompt,
+                schema_cls=cgschema.ReasonScore,
+                extract_schema=lambda r: (r.score, r.reason),
+                extract_json=lambda data: (data["score"], data["reason"]),
+            )
 
     def evaluate(
         self, test_case: ConversationalTestCase
@@ -339,7 +329,7 @@ class ConversationalGEval(BaseConversationalMetric):
             res, cost = self.model.generate_raw_response(
                 prompt, top_logprobs=20
             )
-            self.evaluation_cost += cost
+            self._accrue_cost(cost)
             data = trimAndLoadJson(res.choices[0].message.content, self)
 
             reason = data["reason"]
@@ -356,22 +346,13 @@ class ConversationalGEval(BaseConversationalMetric):
                 return score, reason
         except AttributeError:
             # This catches the case where a_generate_raw_response doesn't exist.
-            if self.using_native_model:
-                res, cost = self.model.generate(
-                    prompt, schema=cgschema.ReasonScore
-                )
-                self.evaluation_cost += cost
-                return res.score, res.reason
-            else:
-                try:
-                    res: cgschema.ReasonScore = self.model.generate(
-                        prompt, schema=cgschema.ReasonScore
-                    )
-                    return res.score, res.reason
-                except TypeError:
-                    res = self.model.generate(prompt)
-                    data = trimAndLoadJson(res, self)
-                    return data["score"], data["reason"]
+            return generate_with_schema_and_extract(
+                metric=self,
+                prompt=prompt,
+                schema_cls=cgschema.ReasonScore,
+                extract_schema=lambda r: (r.score, r.reason),
+                extract_json=lambda data: (data["score"], data["reason"]),
+            )
 
     def generate_weighted_summed_score(
         self, raw_score: int, raw_response: ChatCompletion

deepeval/metrics/dag/dag.py

@@ -62,10 +62,15 @@ class DAGMetric(BaseMetric):
         _in_component: bool = False,
         _log_metric_to_confident: bool = True,
     ) -> float:
+        multimodal = test_case.multimodal
         check_llm_test_case_params(
             test_case,
             extract_required_params(self.dag.root_nodes, self.dag.multiturn),
+            None,
+            None,
             self,
+            self.model,
+            multimodal,
         )
 
         self.evaluation_cost = 0 if self.using_native_model else None
@@ -105,10 +110,15 @@ class DAGMetric(BaseMetric):
         _in_component: bool = False,
         _log_metric_to_confident: bool = True,
     ) -> float:
+        multimodal = test_case.multimodal
         check_llm_test_case_params(
             test_case,
             extract_required_params(self.dag.root_nodes, self.dag.multiturn),
+            None,
+            None,
             self,
+            self.model,
+            multimodal,
         )
 
         self.evaluation_cost = 0 if self.using_native_model else None