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

deepeval/metrics/turn_contextual_relevancy/turn_contextual_relevancy.py

@@ -1,6 +1,6 @@
 from typing import List, Optional, Union, Type, Tuple
 import asyncio
-
+import itertools
 from deepeval.test_case import ConversationalTestCase, TurnParams, Turn
 from deepeval.metrics import BaseConversationalMetric
 from deepeval.utils import (
@@ -12,7 +12,10 @@ from deepeval.metrics.utils import (
     trimAndLoadJson,
     check_conversational_test_case_params,
     get_unit_interactions,
+    get_turns_in_sliding_window,
     initialize_model,
+    generate_with_schema_and_extract,
+    a_generate_with_schema_and_extract,
 )
 from deepeval.models import DeepEvalBaseLLM
 from deepeval.metrics.turn_contextual_relevancy.template import (
@@ -30,6 +33,7 @@ from deepeval.metrics.api import metric_data_manager
 
 class TurnContextualRelevancyMetric(BaseConversationalMetric):
     _required_test_case_params: List[TurnParams] = [
+        TurnParams.ROLE,
         TurnParams.CONTENT,
         TurnParams.RETRIEVAL_CONTEXT,
     ]
@@ -42,6 +46,7 @@ class TurnContextualRelevancyMetric(BaseConversationalMetric):
         async_mode: bool = True,
         strict_mode: bool = False,
         verbose_mode: bool = False,
+        window_size: int = 10,
         evaluation_template: Type[
             TurnContextualRelevancyTemplate
         ] = TurnContextualRelevancyTemplate,
@@ -53,6 +58,7 @@ class TurnContextualRelevancyMetric(BaseConversationalMetric):
         self.async_mode = async_mode
         self.strict_mode = strict_mode
         self.verbose_mode = verbose_mode
+        self.window_size = window_size
         self.evaluation_template = evaluation_template
 
     def measure(
@@ -89,9 +95,19 @@ class TurnContextualRelevancyMetric(BaseConversationalMetric):
                 )
             else:
                 unit_interactions = get_unit_interactions(test_case.turns)
-                scores = self._get_contextual_relevancy_scores(
-                    unit_interactions, multimodal
-                )
+                turns_windows: List[List[Turn]] = [
+                    list(itertools.chain(*window))
+                    for window in get_turns_in_sliding_window(
+                        unit_interactions, self.window_size
+                    )
+                ]
+                scores = []
+                for window in turns_windows:
+                    scores.extend(
+                        self._get_contextual_relevancy_scores(
+                            window, multimodal
+                        )
+                    )
                 self.score = self._calculate_score(scores)
                 self.success = self.score >= self.threshold
                 self.reason = self._generate_reason(scores)
@@ -137,9 +153,25 @@ class TurnContextualRelevancyMetric(BaseConversationalMetric):
             _in_component=_in_component,
         ):
             unit_interactions = get_unit_interactions(test_case.turns)
-            scores = await self._a_get_contextual_relevancy_scores(
-                unit_interactions, multimodal
-            )
+            turns_windows: List[List[Turn]] = [
+                list(itertools.chain(*window))
+                for window in get_turns_in_sliding_window(
+                    unit_interactions, self.window_size
+                )
+            ]
+            scores = []
+            tasks = []
+
+            async def get_individual_scores(window):
+                scores.extend(
+                    await self._a_get_contextual_relevancy_scores(
+                        window, multimodal
+                    )
+                )
+
+            for window in turns_windows:
+                tasks.append(get_individual_scores(window))
+            await asyncio.gather(*tasks)
             self.score = self._calculate_score(scores)
             self.success = self.score >= self.threshold
             self.reason = await self._a_generate_reason(scores)
@@ -160,69 +192,63 @@ class TurnContextualRelevancyMetric(BaseConversationalMetric):
             return self.score
 
     async def _a_get_contextual_relevancy_scores(
-        self, unit_interactions: List[List[Turn]], multimodal: bool
+        self, turns_window: List[Turn], multimodal: bool
     ):
-        async def get_interaction_score(unit_interaction: List[Turn]):
-            user_content = "User Message: "
-            retrieval_context = []
-            for turn in unit_interaction:
-                if turn.role == "user":
-                    user_content += f"\n{turn.content} "
-                else:
-                    retrieval_context.extend(turn.retrieval_context)
+        windows_scores = []
 
-            # Generate verdicts for each retrieval context
-            verdicts = await self._a_generate_verdicts(
-                user_content, retrieval_context, multimodal
-            )
-            score, reason = await self._a_get_interaction_score_and_reason(
-                user_content, verdicts, multimodal
-            )
-            interaction_score = InteractionContextualRelevancyScore(
-                score=score,
-                reason=reason,
-                verdicts=verdicts,
-            )
-            return interaction_score
+        user_content = ""
+        retrieval_context = []
+        for turn in turns_window:
+            if turn.role == "user":
+                user_content += f"\n{turn.content} "
+            else:
+                if turn.retrieval_context is not None:
+                    retrieval_context.extend(turn.retrieval_context)
 
-        final_scores = await asyncio.gather(
-            *[
-                get_interaction_score(unit_interaction)
-                for unit_interaction in unit_interactions
-            ]
+        verdicts = await self._a_generate_verdicts(
+            user_content, retrieval_context, multimodal
+        )
+        score, reason = await self._a_get_interaction_score_and_reason(
+            user_content, verdicts, multimodal
+        )
+        interaction_score = InteractionContextualRelevancyScore(
+            score=score,
+            reason=reason,
+            verdicts=verdicts,
         )
 
-        return final_scores
+        windows_scores.append(interaction_score)
+
+        return windows_scores
 
     def _get_contextual_relevancy_scores(
-        self, unit_interactions: List[List[Turn]], multimodal: bool
+        self, turns_window: List[Turn], multimodal: bool
     ):
-        interaction_scores = []
-
-        for unit_interaction in unit_interactions:
-            user_content = "User Message: "
-            retrieval_context = []
-            for turn in unit_interaction:
-                if turn.role == "user":
-                    user_content += f"\n{turn.content} "
-                else:
+        windows_scores = []
+
+        user_content = ""
+        retrieval_context = []
+        for turn in turns_window:
+            if turn.role == "user":
+                user_content += f"\n{turn.content} "
+            else:
+                if turn.retrieval_context is not None:
                     retrieval_context.extend(turn.retrieval_context)
 
-            # Generate verdicts for each retrieval context
-            verdicts = self._generate_verdicts(
-                user_content, retrieval_context, multimodal
-            )
-            score, reason = self._get_interaction_score_and_reason(
-                user_content, verdicts, multimodal
-            )
-            interaction_score = InteractionContextualRelevancyScore(
-                score=score,
-                reason=reason,
-                verdicts=verdicts,
-            )
-            interaction_scores.append(interaction_score)
+        verdicts = self._generate_verdicts(
+            user_content, retrieval_context, multimodal
+        )
+        score, reason = self._get_interaction_score_and_reason(
+            user_content, verdicts, multimodal
+        )
+        interaction_score = InteractionContextualRelevancyScore(
+            score=score,
+            reason=reason,
+            verdicts=verdicts,
+        )
+        windows_scores.append(interaction_score)
 
-        return interaction_scores
+        return windows_scores
 
     async def _a_generate_verdicts(
         self, input: str, retrieval_context: List[str], multimodal: bool
@@ -240,29 +266,15 @@ class TurnContextualRelevancyMetric(BaseConversationalMetric):
                 multimodal=multimodal,
             )
 
-            if self.using_native_model:
-                res, cost = await self.model.a_generate(
-                    prompt, schema=ContextualRelevancyVerdicts
-                )
-                self.evaluation_cost += cost
-                verdicts.extend([item for item in res.verdicts])
-            else:
-                try:
-                    res: ContextualRelevancyVerdicts = (
-                        await self.model.a_generate(
-                            prompt, schema=ContextualRelevancyVerdicts
-                        )
-                    )
-                    verdicts.extend([item for item in res.verdicts])
-                except TypeError:
-                    res = await self.model.a_generate(prompt)
-                    data = trimAndLoadJson(res, self)
-                    verdicts.extend(
-                        [
-                            ContextualRelevancyVerdict(**item)
-                            for item in data["verdicts"]
-                        ]
-                    )
+            result = await a_generate_with_schema_and_extract(
+                metric=self,
+                prompt=prompt,
+                schema_cls=ContextualRelevancyVerdicts,
+                extract_schema=lambda s: s.verdicts,
+                extract_json=lambda data: data["verdicts"],
+            )
+
+            verdicts.extend(result)
 
         return verdicts
 
@@ -282,27 +294,15 @@ class TurnContextualRelevancyMetric(BaseConversationalMetric):
                 multimodal=multimodal,
             )
 
-            if self.using_native_model:
-                res, cost = self.model.generate(
-                    prompt, schema=ContextualRelevancyVerdicts
-                )
-                self.evaluation_cost += cost
-                verdicts.extend([item for item in res.verdicts])
-            else:
-                try:
-                    res: ContextualRelevancyVerdicts = self.model.generate(
-                        prompt, schema=ContextualRelevancyVerdicts
-                    )
-                    verdicts.extend([item for item in res.verdicts])
-                except TypeError:
-                    res = self.model.generate(prompt)
-                    data = trimAndLoadJson(res, self)
-                    verdicts.extend(
-                        [
-                            ContextualRelevancyVerdict(**item)
-                            for item in data["verdicts"]
-                        ]
-                    )
+            result = generate_with_schema_and_extract(
+                metric=self,
+                prompt=prompt,
+                schema_cls=ContextualRelevancyVerdicts,
+                extract_schema=lambda s: s.verdicts,
+                extract_json=lambda data: data["verdicts"],
+            )
+
+            verdicts.extend(result)
 
         return verdicts
 
@@ -313,7 +313,10 @@ class TurnContextualRelevancyMetric(BaseConversationalMetric):
         multimodal: bool,
     ) -> Tuple[float, str]:
         if len(verdicts) == 0:
-            return 1, None
+            return (
+                1,
+                "There were no retrieval contexts in the given turns to evaluate the contextual relevancy.",
+            )
 
         score = self._calculate_interaction_score(verdicts)
         reason = await self._a_get_interaction_reason(
@@ -332,7 +335,10 @@ class TurnContextualRelevancyMetric(BaseConversationalMetric):
         multimodal: bool,
     ) -> Tuple[float, str]:
         if len(verdicts) == 0:
-            return 1, None
+            return (
+                1,
+                "There were no retrieval contexts in the given turns to evaluate the contextual relevancy.",
+            )
 
         score = self._calculate_interaction_score(verdicts)
         reason = self._get_interaction_reason(
@@ -377,7 +383,6 @@ class TurnContextualRelevancyMetric(BaseConversationalMetric):
             if verdict.verdict.strip().lower() == "yes":
                 relevant_statements.append(verdict.statement)
             else:
-                # Include the reason for irrelevance
                 irrelevant_statements.append(
                     f"{verdict.statement}: {verdict.reason}"
                 )
@@ -390,24 +395,13 @@ class TurnContextualRelevancyMetric(BaseConversationalMetric):
             multimodal=multimodal,
         )
 
-        if self.using_native_model:
-            res, cost = await self.model.a_generate(
-                prompt, schema=ContextualRelevancyScoreReason
-            )
-            self.evaluation_cost += cost
-            return res.reason
-        else:
-            try:
-                res: ContextualRelevancyScoreReason = (
-                    await self.model.a_generate(
-                        prompt, schema=ContextualRelevancyScoreReason
-                    )
-                )
-                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=ContextualRelevancyScoreReason,
+            extract_schema=lambda s: s.reason,
+            extract_json=lambda data: data["reason"],
+        )
 
     def _get_interaction_reason(
         self,
@@ -440,30 +434,21 @@ class TurnContextualRelevancyMetric(BaseConversationalMetric):
             multimodal=multimodal,
         )
 
-        if self.using_native_model:
-            res, cost = self.model.generate(
-                prompt, schema=ContextualRelevancyScoreReason
-            )
-            self.evaluation_cost += cost
-            return res.reason
-        else:
-            try:
-                res: ContextualRelevancyScoreReason = self.model.generate(
-                    prompt, schema=ContextualRelevancyScoreReason
-                )
-                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=ContextualRelevancyScoreReason,
+            extract_schema=lambda s: s.reason,
+            extract_json=lambda data: data["reason"],
+        )
 
     def _get_verbose_steps(
-        self, interaction_scores: List[InteractionContextualRelevancyScore]
+        self, windows_scores: List[InteractionContextualRelevancyScore]
     ):
         steps = []
-        for index, interaction_score in enumerate(interaction_scores):
+        for index, interaction_score in enumerate(windows_scores):
             interaction_steps = [
-                f"Interaction {index + 1} \n",
+                f"Window {index + 1} \n",
                 f"Verdicts: {prettify_list(interaction_score.verdicts)} \n",
                 f"Score: {interaction_score.score} \n",
                 f"Reason: {interaction_score.reason} \n",
@@ -474,6 +459,12 @@ class TurnContextualRelevancyMetric(BaseConversationalMetric):
     def _generate_reason(
         self, scores: List[InteractionContextualRelevancyScore]
     ) -> str:
+        if self.include_reason is False:
+            return None
+
+        if len(scores) == 0:
+            return "There were no retrieval contexts in your turns to evaluate, hence the score is 1"
+
         reasons = []
         for score in scores:
             reasons.append(score.reason)
@@ -482,17 +473,23 @@ class TurnContextualRelevancyMetric(BaseConversationalMetric):
             self.score, self.success, reasons
         )
 
-        if self.using_native_model:
-            res, cost = self.model.generate(prompt)
-            self.evaluation_cost += cost
-            return res
-        else:
-            res = self.model.generate(prompt)
-            return res
+        return generate_with_schema_and_extract(
+            metric=self,
+            prompt=prompt,
+            schema_cls=ContextualRelevancyScoreReason,
+            extract_schema=lambda s: s.reason,
+            extract_json=lambda data: data["reason"],
+        )
 
     async def _a_generate_reason(
         self, scores: List[InteractionContextualRelevancyScore]
     ) -> str:
+        if self.include_reason is False:
+            return None
+
+        if len(scores) == 0:
+            return "There were no retrieval contexts in your turns to evaluate, hence the score is 1"
+
         reasons = []
         for score in scores:
             reasons.append(score.reason)
@@ -501,13 +498,13 @@ class TurnContextualRelevancyMetric(BaseConversationalMetric):
             self.score, self.success, reasons
         )
 
-        if self.using_native_model:
-            res, cost = await self.model.a_generate(prompt)
-            self.evaluation_cost += cost
-            return res
-        else:
-            res = await self.model.a_generate(prompt)
-            return res
+        return await a_generate_with_schema_and_extract(
+            metric=self,
+            prompt=prompt,
+            schema_cls=ContextualRelevancyScoreReason,
+            extract_schema=lambda s: s.reason,
+            extract_json=lambda data: data["reason"],
+        )
 
     def _calculate_score(
         self, scores: List[InteractionContextualRelevancyScore]

deepeval/metrics/turn_faithfulness/schema.py

@@ -25,7 +25,7 @@ class FaithfulnessScoreReason(BaseModel):
 
 class InteractionFaithfulnessScore(BaseModel):
     score: float
-    reason: str
+    reason: Optional[str]
     claims: List[str]
     truths: List[str]
     verdicts: List[FaithfulnessVerdict]

deepeval/metrics/turn_faithfulness/template.py

@@ -187,6 +187,13 @@ class TurnFaithfulnessTemplate:
                 Context:
                 This metric evaluates conversational faithfulness by extracting truths from retrieval context, extracting claims from the assistant's output, and generating verdicts that compare each claim against the truths. Each interaction yields a reason indicating why a verdict failed or succeeded. You are given all those reasons.
 
+                **
+                IMPORTANT: Please make sure to only return in JSON format, with the 'reason' key providing the reason.
+                Example JSON:
+                {{
+                    "reason": "The score is <turn_faithfulness_score> because <your_reason>."
+                }}
+
                 Inputs:
                 - final_score: the averaged score across all interactions.
                 - success: whether the metric passed or failed
@@ -213,6 +220,6 @@ class TurnFaithfulnessTemplate:
 
                 Now give me a final reason that explains why the metric passed or failed. Output ONLY the reason and nothing else.
 
-                The final reason:
+                JSON:
             """
         )