deepeval 3.7.2__py3-none-any.whl → 3.7.4__py3-none-any.whl

deepeval/metrics/multimodal_metrics/multimodal_contextual_relevancy/template.py

@@ -74,13 +74,13 @@ class MultimodalContextualRelevancyTemplate:
                 {{
                     "verdicts": [
                         {{
-                            "verdict": "yes",
                             "statement": "Einstein won the Nobel Prize for his discovery of the photoelectric effect in 1968",
+                            "verdict": "yes"
                         }},
                         {{
-                            "verdict": "no",
                             "statement": "There was a cat.",
-                            "reason": "The retrieval context contained the information 'There was a cat' when it has nothing to do with Einstein's achievements."
+                            "reason": "The retrieval context contained the information 'There was a cat' when it has nothing to do with Einstein's achievements.",
+                            "verdict": "no"
                         }}
                     ]
                 }}

deepeval/metrics/multimodal_metrics/multimodal_faithfulness/template.py

@@ -107,24 +107,24 @@ class MultimodalFaithfulnessTemplate:
                 {{
                     "verdicts": [
                         {{
-                            "verdict": "idk",
-                            "reason": "The claim about Barack Obama is not directly addressed in the retrieval context, and so poses no contradiction."
+                            "reason": "The claim about Barack Obama is not directly addressed in the retrieval context, and so poses no contradiction.",
+                            "verdict": "idk"
                         }},
                         {{
-                            "verdict": "idk",
-                            "reason": "The claim about Zurich being a city in London is incorrect but does not pose a contradiction to the retrieval context."
+                            "reason": "The claim about Zurich being a city in London is incorrect but does not pose a contradiction to the retrieval context.",
+                            "verdict": "idk"
                         }},
                         {{
                             "verdict": "yes"
                         }},
                         {{
-                            "verdict": "no",
-                            "reason": "The actual output claims Einstein won the Nobel Prize in 1969, which is untrue as the retrieval context states it is 1968 instead."
+                            "reason": "The actual output claims Einstein won the Nobel Prize in 1969, which is untrue as the retrieval context states it is 1968 instead.",
+                            "verdict": "no"
                         }},
                         {{
-                            "verdict": "no",
-                            "reason": "The actual output claims Einstein is a German chef, which is not correct as the retrieval context states he was a German scientist instead."
-                        }},
+                            "reason": "The actual output claims Einstein is a German chef, which is not correct as the retrieval context states he was a German scientist instead.",
+                            "verdict": "no"
+                        }}
                     ]  
                 }}
                 ===== END OF EXAMPLE ======

deepeval/metrics/multimodal_metrics/multimodal_g_eval/template.py

@@ -92,8 +92,8 @@ class MultimodalGEvalTemplate:
                 ---
                 **Example JSON:**
                 {{
-                    "score": {score_range[0]},
-                    "reason": "your concise and informative reason here"
+                    "reason": "your concise and informative reason here",
+                    "score": {score_range[0]}
                 }}
 
                 JSON:
@@ -136,8 +136,8 @@ class MultimodalGEvalTemplate:
 
                 Example JSON:
                 {{
-                    "score": 0,
-                    "reason": "The text does not follow the evaluation steps provided."
+                    "reason": "The text does not follow the evaluation steps provided.",
+                    "score": 0
                 }}
                 **
 

deepeval/metrics/non_advice/template.py

@@ -42,8 +42,8 @@ Example JSON:
 {{
     "verdicts": [
         {{
-            "verdict": "yes",
-            "reason": "Contains specific investment advice"
+            "reason": "Contains specific investment advice",
+            "verdict": "yes"
         }}
     ]
 }}

deepeval/metrics/pii_leakage/template.py

@@ -39,8 +39,8 @@ Example JSON:
 {{
     "verdicts": [
         {{
-            "verdict": "yes",
-            "reason": "Contains personal phone number"
+            "reason": "Contains personal phone number",
+            "verdict": "yes"
         }}
     ]
 }}

deepeval/metrics/prompt_alignment/template.py

@@ -26,12 +26,12 @@ Example JSON:
             "verdict": "yes"
         }},
         {{
-            "verdict": "no",
-            "reason": "The LLM corrected the user when the user used the wrong grammar in asking about the number of stars in the sky."
+            "reason": "The LLM corrected the user when the user used the wrong grammar in asking about the number of stars in the sky.",
+            "verdict": "no"
         }},
         {{
-            "verdict": "no",
-            "reason": "The LLM only made 'HEY THERE' uppercase, which does not follow the instruction of making everything uppercase completely."
+            "reason": "The LLM only made 'HEY THERE' uppercase, which does not follow the instruction of making everything uppercase completely.",
+            "verdict": "no"
         }}
     ]  
 }}

deepeval/metrics/role_violation/template.py

@@ -39,8 +39,8 @@ Example JSON:
 {{
     "verdicts": [
         {{
-            "verdict": "yes",
-            "reason": "AI is pretending to be human"
+            "reason": "AI is pretending to be human",
+            "verdict": "yes"
         }}
     ]
 }}

deepeval/metrics/step_efficiency/step_efficiency.py

@@ -231,4 +231,4 @@ class StepEfficiencyMetric(BaseMetric):
 
     @property
     def __name__(self):
-        return "Execution Efficiency"
+        return "Step Efficiency"

deepeval/metrics/toxicity/template.py

@@ -71,15 +71,15 @@ Example JSON:
 {{
     "verdicts": [
         {{
-            "verdict": "yes",
-            "reason": "The actual output uses words like "garbage," "tasteless weeds," etc., ridicules vegan food and those who choose it, and attacks people who enjoy vegan food, calling them "pretentious losers. Furthermore, it dismisses the entire concept of vegan cuisine with a sweeping generalization, making it seem worthless."
+            "reason": "The actual output uses words like "garbage," "tasteless weeds," etc., ridicules vegan food and those who choose it, and attacks people who enjoy vegan food, calling them "pretentious losers. Furthermore, it dismisses the entire concept of vegan cuisine with a sweeping generalization, making it seem worthless.",
+            "verdict": "yes"
         }},
         {{
             "verdict": "no"
         }},
         {{
-            "verdict": "yes":
-            "reason": "'Sarah always meant well' sounds positive but are undermined by the surrounding criticism such as 'can't help but sign', which can be considered a personal attack."
+            "reason": "'Sarah always meant well' sounds positive but are undermined by the surrounding criticism such as 'can't help but sign', which can be considered a personal attack.",
+            "verdict": "yes"
         }}
     ]  
 }}

deepeval/metrics/turn_relevancy/template.py

@@ -33,8 +33,8 @@ Example Messages:
 
 Example JSON:
 {{
-    "verdict": "no",
-    "reason": "The LLM responded 'isn't it a nice day today' to a message that asked about how to treat a sore throat, which is completely irrelevant."
+    "reason": "The LLM responded 'isn't it a nice day today' to a message that asked about how to treat a sore throat, which is completely irrelevant.",
+    "verdict": "no"
 }}
 ===== END OF EXAMPLE ======
 You MUST ONLY provide a verdict for the LAST message on the list but MUST USE context from the previous messages.

deepeval/metrics/utils.py

@@ -25,6 +25,7 @@ from deepeval.models import (
     MultimodalOpenAIModel,
     MultimodalGeminiModel,
     MultimodalOllamaModel,
+    MultimodalAzureOpenAIMLLMModel,
     AmazonBedrockModel,
     LiteLLMModel,
     KimiModel,
@@ -514,6 +515,8 @@ def initialize_multimodal_model(
         return MultimodalGeminiModel(), True
     if should_use_ollama_model():
         return MultimodalOllamaModel(), True
+    elif should_use_azure_openai():
+        return MultimodalAzureOpenAIMLLMModel(model_name=model), True
     elif isinstance(model, str) or model is None:
         return MultimodalOpenAIModel(model=model), True
     raise TypeError(

deepeval/models/__init__.py

@@ -21,6 +21,7 @@ from deepeval.models.mlllms import (
     MultimodalOpenAIModel,
     MultimodalOllamaModel,
     MultimodalGeminiModel,
+    MultimodalAzureOpenAIMLLMModel,
 )
 from deepeval.models.embedding_models import (
     OpenAIEmbeddingModel,
@@ -48,6 +49,7 @@ __all__ = [
     "MultimodalOpenAIModel",
     "MultimodalOllamaModel",
     "MultimodalGeminiModel",
+    "MultimodalAzureOpenAIMLLMModel",
     "OpenAIEmbeddingModel",
     "AzureOpenAIEmbeddingModel",
     "LocalEmbeddingModel",

deepeval/models/embedding_models/azure_embedding_model.py

@@ -1,16 +1,15 @@
 from typing import Dict, List, Optional
 from openai import AzureOpenAI, AsyncAzureOpenAI
-from deepeval.key_handler import (
-    EmbeddingKeyValues,
-    ModelKeyValues,
-    KEY_FILE_HANDLER,
-)
+from pydantic import SecretStr
+
+from deepeval.config.settings import get_settings
 from deepeval.models import DeepEvalBaseEmbeddingModel
 from deepeval.models.retry_policy import (
     create_retry_decorator,
     sdk_retries_for,
 )
 from deepeval.constants import ProviderSlug as PS
+from deepeval.models.utils import require_secret_api_key
 
 
 retry_azure = create_retry_decorator(PS.AZURE)
@@ -27,18 +26,25 @@ class AzureOpenAIEmbeddingModel(DeepEvalBaseEmbeddingModel):
         generation_kwargs: Optional[Dict] = None,
         **client_kwargs,
     ):
-        self.openai_api_key = openai_api_key or KEY_FILE_HANDLER.fetch_data(
-            ModelKeyValues.AZURE_OPENAI_API_KEY
-        )
+        settings = get_settings()
+
+        if openai_api_key is not None:
+            # keep it secret, keep it safe from serializings, logging and alike
+            self.openai_api_key: SecretStr | None = SecretStr(openai_api_key)
+        else:
+            self.openai_api_key = settings.AZURE_OPENAI_API_KEY
+
         self.openai_api_version = (
-            openai_api_version
-            or KEY_FILE_HANDLER.fetch_data(ModelKeyValues.OPENAI_API_VERSION)
+            openai_api_version or settings.OPENAI_API_VERSION
         )
-        self.azure_endpoint = azure_endpoint or KEY_FILE_HANDLER.fetch_data(
-            ModelKeyValues.AZURE_OPENAI_ENDPOINT
+        self.azure_endpoint = (
+            azure_endpoint
+            or settings.AZURE_OPENAI_ENDPOINT
+            and str(settings.AZURE_OPENAI_ENDPOINT)
         )
-        self.azure_deployment = azure_deployment or KEY_FILE_HANDLER.fetch_data(
-            EmbeddingKeyValues.AZURE_EMBEDDING_DEPLOYMENT_NAME
+
+        self.azure_deployment = (
+            azure_deployment or settings.AZURE_EMBEDDING_DEPLOYMENT_NAME
         )
         self.client_kwargs = client_kwargs or {}
         self.model_name = model or self.azure_deployment
@@ -86,12 +92,19 @@ class AzureOpenAIEmbeddingModel(DeepEvalBaseEmbeddingModel):
         return self._build_client(AsyncAzureOpenAI)
 
     def _build_client(self, cls):
+        api_key = require_secret_api_key(
+            self.openai_api_key,
+            provider_label="AzureOpenAI",
+            env_var_name="AZURE_OPENAI_API_KEY",
+            param_hint="`openai_api_key` to AzureOpenAIEmbeddingModel(...)",
+        )
+
         client_kwargs = self.client_kwargs.copy()
         if not sdk_retries_for(PS.AZURE):
             client_kwargs["max_retries"] = 0
 
         client_init_kwargs = dict(
-            api_key=self.openai_api_key,
+            api_key=api_key,
             api_version=self.openai_api_version,
             azure_endpoint=self.azure_endpoint,
             azure_deployment=self.azure_deployment,

deepeval/models/embedding_models/local_embedding_model.py

@@ -1,7 +1,9 @@
 from openai import OpenAI, AsyncOpenAI
 from typing import Dict, List, Optional
+from pydantic import SecretStr
 
-from deepeval.key_handler import EmbeddingKeyValues, KEY_FILE_HANDLER
+from deepeval.config.settings import get_settings
+from deepeval.models.utils import require_secret_api_key
 from deepeval.models import DeepEvalBaseEmbeddingModel
 from deepeval.models.retry_policy import (
     create_retry_decorator,
@@ -23,15 +25,19 @@ class LocalEmbeddingModel(DeepEvalBaseEmbeddingModel):
         generation_kwargs: Optional[Dict] = None,
         **client_kwargs,
     ):
-        self.api_key = api_key or KEY_FILE_HANDLER.fetch_data(
-            EmbeddingKeyValues.LOCAL_EMBEDDING_API_KEY
-        )
-        self.base_url = base_url or KEY_FILE_HANDLER.fetch_data(
-            EmbeddingKeyValues.LOCAL_EMBEDDING_BASE_URL
-        )
-        self.model_name = model or KEY_FILE_HANDLER.fetch_data(
-            EmbeddingKeyValues.LOCAL_EMBEDDING_MODEL_NAME
+        settings = get_settings()
+        if api_key is not None:
+            # keep it secret, keep it safe from serializings, logging and alike
+            self.api_key: SecretStr | None = SecretStr(api_key)
+        else:
+            self.api_key = get_settings().LOCAL_EMBEDDING_API_KEY
+
+        self.base_url = (
+            base_url
+            or settings.LOCAL_EMBEDDING_BASE_URL
+            and str(settings.LOCAL_EMBEDDING_BASE_URL)
         )
+        self.model_name = model or settings.LOCAL_EMBEDDING_MODEL_NAME
         self.client_kwargs = client_kwargs or {}
         self.generation_kwargs = generation_kwargs or {}
         super().__init__(self.model_name)
@@ -81,12 +87,19 @@ class LocalEmbeddingModel(DeepEvalBaseEmbeddingModel):
         return self._build_client(AsyncOpenAI)
 
     def _build_client(self, cls):
+        api_key = require_secret_api_key(
+            self.api_key,
+            provider_label="OpenAI",
+            env_var_name="LOCAL_EMBEDDING_API_KEY",
+            param_hint="`api_key` to LocalEmbeddingModel(...)",
+        )
+
         client_kwargs = self.client_kwargs.copy()
         if not sdk_retries_for(PS.LOCAL):
             client_kwargs["max_retries"] = 0
 
         client_init_kwargs = dict(
-            api_key=self.api_key,
+            api_key=api_key,
             base_url=self.base_url,
             **client_kwargs,
         )

deepeval/models/embedding_models/ollama_embedding_model.py

@@ -1,7 +1,7 @@
 from ollama import Client, AsyncClient
 from typing import List, Optional, Dict
 
-from deepeval.key_handler import EmbeddingKeyValues, KEY_FILE_HANDLER
+from deepeval.config.settings import get_settings
 from deepeval.models import DeepEvalBaseEmbeddingModel
 from deepeval.models.retry_policy import (
     create_retry_decorator,
@@ -20,12 +20,14 @@ class OllamaEmbeddingModel(DeepEvalBaseEmbeddingModel):
         generation_kwargs: Optional[Dict] = None,
         **client_kwargs,
     ):
-        self.host = host or KEY_FILE_HANDLER.fetch_data(
-            EmbeddingKeyValues.LOCAL_EMBEDDING_BASE_URL
-        )
-        self.model_name = model or KEY_FILE_HANDLER.fetch_data(
-            EmbeddingKeyValues.LOCAL_EMBEDDING_MODEL_NAME
+        settings = get_settings()
+
+        self.host = (
+            host
+            or settings.LOCAL_EMBEDDING_BASE_URL
+            and str(settings.LOCAL_EMBEDDING_BASE_URL)
         )
+        self.model_name = model or settings.LOCAL_EMBEDDING_MODEL_NAME
         self.client_kwargs = client_kwargs or {}
         self.generation_kwargs = generation_kwargs or {}
         super().__init__(self.model_name)

deepeval/models/embedding_models/openai_embedding_model.py

@@ -1,5 +1,9 @@
 from typing import Dict, Optional, List
 from openai import OpenAI, AsyncOpenAI
+from pydantic import SecretStr
+
+from deepeval.config.settings import get_settings
+from deepeval.models.utils import require_secret_api_key
 from deepeval.models import DeepEvalBaseEmbeddingModel
 from deepeval.models.retry_policy import (
     create_retry_decorator,
@@ -27,7 +31,12 @@ class OpenAIEmbeddingModel(DeepEvalBaseEmbeddingModel):
         generation_kwargs: Optional[Dict] = None,
         **client_kwargs,
     ):
-        self.openai_api_key = openai_api_key
+        if openai_api_key is not None:
+            # keep it secret, keep it safe from serializings, logging and alike
+            self.openai_api_key: SecretStr | None = SecretStr(openai_api_key)
+        else:
+            self.openai_api_key = get_settings().OPENAI_API_KEY
+
         self.model_name = model if model else default_openai_embedding_model
         if self.model_name not in valid_openai_embedding_models:
             raise ValueError(
@@ -81,12 +90,19 @@ class OpenAIEmbeddingModel(DeepEvalBaseEmbeddingModel):
         return self._build_client(AsyncOpenAI)
 
     def _build_client(self, cls):
+        openai_api_key = require_secret_api_key(
+            self.openai_api_key,
+            provider_label="OpenAI",
+            env_var_name="OPENAI_API_KEY",
+            param_hint="`openai_api_key` to OpenAIEmbeddingModel(...)",
+        )
+
         client_kwargs = self.client_kwargs.copy()
         if not sdk_retries_for(PS.OPENAI):
             client_kwargs["max_retries"] = 0
 
         client_init_kwargs = dict(
-            api_key=self.openai_api_key,
+            api_key=openai_api_key,
             **client_kwargs,
         )
         try:

deepeval/models/llms/anthropic_model.py

@@ -2,7 +2,7 @@ import warnings
 
 from typing import Optional, Tuple, Union, Dict
 from anthropic import Anthropic, AsyncAnthropic
-from pydantic import BaseModel
+from pydantic import BaseModel, SecretStr
 
 from deepeval.models import DeepEvalBaseLLM
 from deepeval.models.llms.utils import trim_and_load_json
@@ -10,7 +10,7 @@ from deepeval.models.retry_policy import (
     create_retry_decorator,
     sdk_retries_for,
 )
-from deepeval.models.utils import parse_model_name
+from deepeval.models.utils import parse_model_name, require_secret_api_key
 from deepeval.config.settings import get_settings
 from deepeval.constants import ProviderSlug as PS
 
@@ -41,7 +41,14 @@ class AnthropicModel(DeepEvalBaseLLM):
         **kwargs,
     ):
         model_name = parse_model_name(model)
-        self._anthropic_api_key = _anthropic_api_key
+
+        if _anthropic_api_key is not None:
+            # keep it secret, keep it safe from serializings, logging and alike
+            self._anthropic_api_key: SecretStr | None = SecretStr(
+                _anthropic_api_key
+            )
+        else:
+            self._anthropic_api_key = get_settings().ANTHROPIC_API_KEY
 
         if temperature < 0:
             raise ValueError("Temperature must be >= 0.")
@@ -155,9 +162,14 @@ class AnthropicModel(DeepEvalBaseLLM):
         return kwargs
 
     def _build_client(self, cls):
-        settings = get_settings()
+        api_key = require_secret_api_key(
+            self._anthropic_api_key,
+            provider_label="Anthropic",
+            env_var_name="ANTHROPIC_API_KEY",
+            param_hint="`_anthropic_api_key` to AnthropicModel(...)",
+        )
         kw = dict(
-            api_key=settings.ANTHROPIC_API_KEY or self._anthropic_api_key,
+            api_key=api_key,
             **self._client_kwargs(),
         )
         try:

deepeval/models/llms/azure_model.py

@@ -1,10 +1,10 @@
 from openai.types.chat.chat_completion import ChatCompletion
 from openai import AzureOpenAI, AsyncAzureOpenAI
 from typing import Optional, Tuple, Union, Dict
-from pydantic import BaseModel
+from pydantic import BaseModel, SecretStr
 
+from deepeval.config.settings import get_settings
 from deepeval.models import DeepEvalBaseLLM
-from deepeval.key_handler import ModelKeyValues, KEY_FILE_HANDLER
 from deepeval.models.llms.openai_model import (
     structured_outputs_models,
     json_mode_models,
@@ -16,7 +16,7 @@ from deepeval.models.retry_policy import (
 )
 
 from deepeval.models.llms.utils import trim_and_load_json
-from deepeval.models.utils import parse_model_name
+from deepeval.models.utils import parse_model_name, require_secret_api_key
 from deepeval.constants import ProviderSlug as PS
 
 
@@ -35,24 +35,29 @@ class AzureOpenAIModel(DeepEvalBaseLLM):
         generation_kwargs: Optional[Dict] = None,
         **kwargs,
     ):
+        settings = get_settings()
+
         # fetch Azure deployment parameters
-        model_name = model_name or KEY_FILE_HANDLER.fetch_data(
-            ModelKeyValues.AZURE_MODEL_NAME
-        )
-        self.deployment_name = deployment_name or KEY_FILE_HANDLER.fetch_data(
-            ModelKeyValues.AZURE_DEPLOYMENT_NAME
-        )
-        self.azure_openai_api_key = (
-            azure_openai_api_key
-            or KEY_FILE_HANDLER.fetch_data(ModelKeyValues.AZURE_OPENAI_API_KEY)
-        )
+        model_name = model_name or settings.AZURE_MODEL_NAME
+        self.deployment_name = deployment_name or settings.AZURE_DEPLOYMENT_NAME
+
+        if azure_openai_api_key is not None:
+            # keep it secret, keep it safe from serializings, logging and alike
+            self.azure_openai_api_key: SecretStr | None = SecretStr(
+                azure_openai_api_key
+            )
+        else:
+            self.azure_openai_api_key = settings.AZURE_OPENAI_API_KEY
+
         self.openai_api_version = (
-            openai_api_version
-            or KEY_FILE_HANDLER.fetch_data(ModelKeyValues.OPENAI_API_VERSION)
+            openai_api_version or settings.OPENAI_API_VERSION
         )
-        self.azure_endpoint = azure_endpoint or KEY_FILE_HANDLER.fetch_data(
-            ModelKeyValues.AZURE_OPENAI_ENDPOINT
+        self.azure_endpoint = (
+            azure_endpoint
+            or settings.AZURE_OPENAI_ENDPOINT
+            and str(settings.AZURE_OPENAI_ENDPOINT)
         )
+
         if temperature < 0:
             raise ValueError("Temperature must be >= 0.")
         self.temperature = temperature
@@ -270,8 +275,15 @@ class AzureOpenAIModel(DeepEvalBaseLLM):
         return kwargs
 
     def _build_client(self, cls):
+        api_key = require_secret_api_key(
+            self.azure_openai_api_key,
+            provider_label="AzureOpenAI",
+            env_var_name="AZURE_OPENAI_API_KEY",
+            param_hint="`azure_openai_api_key` to AzureOpenAIModel(...)",
+        )
+
         kw = dict(
-            api_key=self.azure_openai_api_key,
+            api_key=api_key,
             api_version=self.openai_api_version,
             azure_endpoint=self.azure_endpoint,
             azure_deployment=self.deployment_name,

deepeval/models/llms/deepseek_model.py

@@ -1,9 +1,10 @@
 from typing import Optional, Tuple, Union, Dict
 from openai import OpenAI, AsyncOpenAI
-from pydantic import BaseModel
+from pydantic import BaseModel, SecretStr
 
-from deepeval.key_handler import ModelKeyValues, KEY_FILE_HANDLER
+from deepeval.config.settings import get_settings
 from deepeval.models.llms.utils import trim_and_load_json
+from deepeval.models.utils import require_secret_api_key
 from deepeval.models import DeepEvalBaseLLM
 from deepeval.models.retry_policy import (
     create_retry_decorator,
@@ -36,25 +37,27 @@ class DeepSeekModel(DeepEvalBaseLLM):
         generation_kwargs: Optional[Dict] = None,
         **kwargs,
     ):
-        model_name = model or KEY_FILE_HANDLER.fetch_data(
-            ModelKeyValues.DEEPSEEK_MODEL_NAME
-        )
+        settings = get_settings()
+
+        model_name = model or settings.DEEPSEEK_MODEL_NAME
         if model_name not in model_pricing:
             raise ValueError(
                 f"Invalid model. Available DeepSeek models: {', '.join(model_pricing.keys())}"
             )
-        temperature_from_key = KEY_FILE_HANDLER.fetch_data(
-            ModelKeyValues.TEMPERATURE
-        )
+        temperature_from_key = settings.TEMPERATURE
         if temperature_from_key is None:
             self.temperature = temperature
         else:
             self.temperature = float(temperature_from_key)
         if self.temperature < 0:
             raise ValueError("Temperature must be >= 0.")
-        self.api_key = api_key or KEY_FILE_HANDLER.fetch_data(
-            ModelKeyValues.DEEPSEEK_API_KEY
-        )
+
+        if api_key is not None:
+            # keep it secret, keep it safe from serializings, logging and alike
+            self.api_key: SecretStr | None = SecretStr(api_key)
+        else:
+            self.api_key = settings.DEEPSEEK_API_KEY
+
         self.base_url = "https://api.deepseek.com"
         self.kwargs = kwargs
         self.generation_kwargs = generation_kwargs or {}
@@ -167,8 +170,15 @@ class DeepSeekModel(DeepEvalBaseLLM):
         return kwargs
 
     def _build_client(self, cls):
+        api_key = require_secret_api_key(
+            self.api_key,
+            provider_label="DeepSeek",
+            env_var_name="DEEPSEEK_API_KEY",
+            param_hint="`api_key` to DeepSeekModel(...)",
+        )
+
         kw = dict(
-            api_key=self.api_key,
+            api_key=api_key,
             base_url=self.base_url,
             **self._client_kwargs(),
         )