deepeval 3.5.2__py3-none-any.whl → 3.5.4__py3-none-any.whl

deepeval/models/llms/amazon_bedrock_model.py

@@ -1,10 +1,16 @@
+import asyncio
+
 from typing import Optional, Tuple, Union, Dict
 from contextlib import AsyncExitStack
 from pydantic import BaseModel
-import asyncio
 
+from deepeval.models.retry_policy import (
+    create_retry_decorator,
+    sdk_retries_for,
+)
 from deepeval.models import DeepEvalBaseLLM
 from deepeval.models.llms.utils import trim_and_load_json
+from deepeval.constants import ProviderSlug as PS
 
 # check aiobotocore availability
 try:
@@ -15,6 +21,9 @@ try:
 except ImportError:
     aiobotocore_available = False
 
+# define retry policy
+retry_bedrock = create_retry_decorator(PS.BEDROCK)
+
 
 def _check_aiobotocore_available():
     if not aiobotocore_available:
@@ -53,11 +62,11 @@ class AmazonBedrockModel(DeepEvalBaseLLM):
 
         # prepare aiobotocore session, config, and async exit stack
         self._session = get_session()
-        self._config = Config(retries={"max_attempts": 5, "mode": "adaptive"})
         self._exit_stack = AsyncExitStack()
         self.kwargs = kwargs
         self.generation_kwargs = generation_kwargs or {}
         self._client = None
+        self._sdk_retry_mode: Optional[bool] = None
 
     ###############################################
     # Generate functions
@@ -68,6 +77,7 @@ class AmazonBedrockModel(DeepEvalBaseLLM):
     ) -> Tuple[Union[str, Dict], float]:
         return asyncio.run(self.a_generate(prompt, schema))
 
+    @retry_bedrock
     async def a_generate(
         self, prompt: str, schema: Optional[BaseModel] = None
     ) -> Tuple[Union[str, Dict], float]:
@@ -94,16 +104,33 @@ class AmazonBedrockModel(DeepEvalBaseLLM):
     ###############################################
 
     async def _ensure_client(self):
-        if self._client is None:
+        use_sdk = sdk_retries_for(PS.BEDROCK)
+
+        # only rebuild if client is missing or the sdk retry mode changes
+        if self._client is None or self._sdk_retry_mode != use_sdk:
+            # Close any previous
+            if self._client is not None:
+                await self._exit_stack.aclose()
+                self._client = None
+
+            # create retry config for botocore
+            retries_config = {"max_attempts": (5 if use_sdk else 1)}
+            if use_sdk:
+                retries_config["mode"] = "adaptive"
+
+            config = Config(retries=retries_config)
+
             cm = self._session.create_client(
                 "bedrock-runtime",
                 region_name=self.region_name,
                 aws_access_key_id=self.aws_access_key_id,
                 aws_secret_access_key=self.aws_secret_access_key,
-                config=self._config,
+                config=config,
                 **self.kwargs,
             )
             self._client = await self._exit_stack.enter_async_context(cm)
+            self._sdk_retry_mode = use_sdk
+
         return self._client
 
     async def close(self):

deepeval/models/llms/anthropic_model.py

@@ -1,12 +1,22 @@
+import warnings
+
 from typing import Optional, Tuple, Union, Dict
 from anthropic import Anthropic, AsyncAnthropic
 from pydantic import BaseModel
-import os
-import warnings
 
 from deepeval.models import DeepEvalBaseLLM
 from deepeval.models.llms.utils import trim_and_load_json
+from deepeval.models.retry_policy import (
+    create_retry_decorator,
+    sdk_retries_for,
+)
 from deepeval.models.utils import parse_model_name
+from deepeval.config.settings import get_settings
+from deepeval.constants import ProviderSlug as PS
+
+
+# consistent retry rules
+retry_anthropic = create_retry_decorator(PS.ANTHROPIC)
 
 model_pricing = {
     "claude-opus-4-20250514": {"input": 15.00 / 1e6, "output": 75.00 / 1e6},
@@ -45,6 +55,7 @@ class AnthropicModel(DeepEvalBaseLLM):
     # Generate functions
     ###############################################
 
+    @retry_anthropic
     def generate(
         self, prompt: str, schema: Optional[BaseModel] = None
     ) -> Tuple[Union[str, Dict], float]:
@@ -70,6 +81,7 @@ class AnthropicModel(DeepEvalBaseLLM):
             json_output = trim_and_load_json(message.content[0].text)
             return schema.model_validate(json_output), cost
 
+    @retry_anthropic
     async def a_generate(
         self, prompt: str, schema: Optional[BaseModel] = None
     ) -> Tuple[str, float]:
@@ -128,17 +140,31 @@ class AnthropicModel(DeepEvalBaseLLM):
 
     def load_model(self, async_mode: bool = False):
         if not async_mode:
-            return Anthropic(
-                api_key=os.environ.get("ANTHROPIC_API_KEY")
-                or self._anthropic_api_key,
-                **self.kwargs,
-            )
-        else:
-            return AsyncAnthropic(
-                api_key=os.environ.get("ANTHROPIC_API_KEY")
-                or self._anthropic_api_key,
-                **self.kwargs,
-            )
+            return self._build_client(Anthropic)
+        return self._build_client(AsyncAnthropic)
 
     def get_model_name(self):
         return f"{self.model_name}"
+
+    def _client_kwargs(self) -> Dict:
+        kwargs = dict(self.kwargs or {})
+        # If we are managing retries with Tenacity, force SDK retries off to avoid double retries.
+        # if the user opts into SDK retries via DEEPEVAL_SDK_RETRY_PROVIDERS, then honor their max_retries.
+        if not sdk_retries_for(PS.ANTHROPIC):
+            kwargs["max_retries"] = 0
+        return kwargs
+
+    def _build_client(self, cls):
+        settings = get_settings()
+        kw = dict(
+            api_key=settings.ANTHROPIC_API_KEY or self._anthropic_api_key,
+            **self._client_kwargs(),
+        )
+        try:
+            return cls(**kw)
+        except TypeError as e:
+            # in case older SDKs don’t accept max_retries, drop it and retry
+            if "max_retries" in str(e):
+                kw.pop("max_retries", None)
+                return cls(**kw)
+            raise

deepeval/models/llms/azure_model.py

@@ -1,9 +1,7 @@
-from tenacity import retry, retry_if_exception_type, wait_exponential_jitter
 from openai.types.chat.chat_completion import ChatCompletion
 from openai import AzureOpenAI, AsyncAzureOpenAI
 from typing import Optional, Tuple, Union, Dict
 from pydantic import BaseModel
-import openai
 
 from deepeval.models import DeepEvalBaseLLM
 from deepeval.key_handler import ModelKeyValues, KEY_FILE_HANDLER
@@ -11,17 +9,18 @@ from deepeval.models.llms.openai_model import (
     structured_outputs_models,
     json_mode_models,
     model_pricing,
-    log_retry_error,
 )
+from deepeval.models.retry_policy import (
+    create_retry_decorator,
+    sdk_retries_for,
+)
+
 from deepeval.models.llms.utils import trim_and_load_json
 from deepeval.models.utils import parse_model_name
+from deepeval.constants import ProviderSlug as PS
 
-retryable_exceptions = (
-    openai.RateLimitError,
-    openai.APIConnectionError,
-    openai.APITimeoutError,
-    openai.LengthFinishReasonError,
-)
+
+retry_azure = create_retry_decorator(PS.AZURE)
 
 
 class AzureOpenAIModel(DeepEvalBaseLLM):
@@ -67,11 +66,7 @@ class AzureOpenAIModel(DeepEvalBaseLLM):
     # Other generate functions
     ###############################################
 
-    @retry(
-        wait=wait_exponential_jitter(initial=1, exp_base=2, jitter=2, max=10),
-        retry=retry_if_exception_type(openai.RateLimitError),
-        after=log_retry_error,
-    )
+    @retry_azure
     def generate(
         self, prompt: str, schema: Optional[BaseModel] = None
     ) -> Tuple[Union[str, Dict], float]:
@@ -130,11 +125,7 @@ class AzureOpenAIModel(DeepEvalBaseLLM):
         else:
             return output, cost
 
-    @retry(
-        wait=wait_exponential_jitter(initial=1, exp_base=2, jitter=2, max=10),
-        retry=retry_if_exception_type(openai.RateLimitError),
-        after=log_retry_error,
-    )
+    @retry_azure
     async def a_generate(
         self, prompt: str, schema: Optional[BaseModel] = None
     ) -> Tuple[Union[str, BaseModel], float]:
@@ -199,11 +190,7 @@ class AzureOpenAIModel(DeepEvalBaseLLM):
     # Other generate functions
     ###############################################
 
-    @retry(
-        wait=wait_exponential_jitter(initial=1, exp_base=2, jitter=2, max=10),
-        retry=retry_if_exception_type(retryable_exceptions),
-        after=log_retry_error,
-    )
+    @retry_azure
     def generate_raw_response(
         self,
         prompt: str,
@@ -226,11 +213,7 @@ class AzureOpenAIModel(DeepEvalBaseLLM):
 
         return completion, cost
 
-    @retry(
-        wait=wait_exponential_jitter(initial=1, exp_base=2, jitter=2, max=10),
-        retry=retry_if_exception_type(retryable_exceptions),
-        after=log_retry_error,
-    )
+    @retry_azure
     async def a_generate_raw_response(
         self,
         prompt: str,
@@ -272,17 +255,33 @@ class AzureOpenAIModel(DeepEvalBaseLLM):
 
     def load_model(self, async_mode: bool = False):
         if not async_mode:
-            return AzureOpenAI(
-                api_key=self.azure_openai_api_key,
-                api_version=self.openai_api_version,
-                azure_endpoint=self.azure_endpoint,
-                azure_deployment=self.deployment_name,
-                **self.kwargs,  # ← Keep this for client initialization
-            )
-        return AsyncAzureOpenAI(
+            return self._build_client(AzureOpenAI)
+        return self._build_client(AsyncAzureOpenAI)
+
+    def _client_kwargs(self) -> Dict:
+        """
+        If Tenacity is managing retries, force OpenAI SDK retries off to avoid double retries.
+        If the user opts into SDK retries for 'azure' via DEEPEVAL_SDK_RETRY_PROVIDERS,
+        leave their retry settings as is.
+        """
+        kwargs = dict(self.kwargs or {})
+        if not sdk_retries_for(PS.AZURE):
+            kwargs["max_retries"] = 0
+        return kwargs
+
+    def _build_client(self, cls):
+        kw = dict(
             api_key=self.azure_openai_api_key,
             api_version=self.openai_api_version,
             azure_endpoint=self.azure_endpoint,
             azure_deployment=self.deployment_name,
-            **self.kwargs,  # ← Keep this for client initialization
+            **self._client_kwargs(),
         )
+        try:
+            return cls(**kw)
+        except TypeError as e:
+            # older OpenAI SDKs may not accept max_retries, in that case remove and retry once
+            if "max_retries" in str(e):
+                kw.pop("max_retries", None)
+                return cls(**kw)
+            raise

deepeval/models/llms/deepseek_model.py

@@ -5,6 +5,15 @@ from pydantic import BaseModel
 from deepeval.key_handler import ModelKeyValues, KEY_FILE_HANDLER
 from deepeval.models.llms.utils import trim_and_load_json
 from deepeval.models import DeepEvalBaseLLM
+from deepeval.models.retry_policy import (
+    create_retry_decorator,
+    sdk_retries_for,
+)
+from deepeval.constants import ProviderSlug as PS
+
+
+# consistent retry rules
+retry_deepseek = create_retry_decorator(PS.DEEPSEEK)
 
 model_pricing = {
     "deepseek-chat": {
@@ -55,6 +64,7 @@ class DeepSeekModel(DeepEvalBaseLLM):
     # Other generate functions
     ###############################################
 
+    @retry_deepseek
     def generate(
         self, prompt: str, schema: Optional[BaseModel] = None
     ) -> Tuple[Union[str, Dict], float]:
@@ -88,6 +98,7 @@ class DeepSeekModel(DeepEvalBaseLLM):
             )
             return output, cost
 
+    @retry_deepseek
     async def a_generate(
         self, prompt: str, schema: Optional[BaseModel] = None
     ) -> Tuple[Union[str, Dict], float]:
@@ -141,13 +152,31 @@ class DeepSeekModel(DeepEvalBaseLLM):
 
     def load_model(self, async_mode: bool = False):
         if not async_mode:
-            return OpenAI(
-                api_key=self.api_key, base_url=self.base_url, **self.kwargs
-            )
-        else:
-            return AsyncOpenAI(
-                api_key=self.api_key, base_url=self.base_url, **self.kwargs
-            )
+            return self._build_client(OpenAI)
+        return self._build_client(AsyncOpenAI)
 
     def get_model_name(self):
         return f"{self.model_name}"
+
+    def _client_kwargs(self) -> Dict:
+        kwargs = dict(self.kwargs or {})
+        # if we are managing retries with Tenacity, force SDK retries off to avoid double retries.
+        # if the user opts into SDK retries for "deepseek" via DEEPEVAL_SDK_RETRY_PROVIDERS, honor it.
+        if not sdk_retries_for(PS.DEEPSEEK):
+            kwargs["max_retries"] = 0
+        return kwargs
+
+    def _build_client(self, cls):
+        kw = dict(
+            api_key=self.api_key,
+            base_url=self.base_url,
+            **self._client_kwargs(),
+        )
+        try:
+            return cls(**kw)
+        except TypeError as e:
+            # In case an older OpenAI client doesn’t accept max_retries, drop it and retry.
+            if "max_retries" in str(e):
+                kw.pop("max_retries", None)
+                return cls(**kw)
+            raise

deepeval/models/llms/gemini_model.py

@@ -3,11 +3,19 @@ from google.genai import types
 from typing import Optional, Dict
 from google import genai
 
+from deepeval.models.retry_policy import (
+    create_retry_decorator,
+)
 from deepeval.key_handler import ModelKeyValues, KEY_FILE_HANDLER
 from deepeval.models.base_model import DeepEvalBaseLLM
+from deepeval.constants import ProviderSlug as PS
+
 
 default_gemini_model = "gemini-1.5-pro"
 
+# consistent retry rules
+retry_gemini = create_retry_decorator(PS.GOOGLE)
+
 
 class GeminiModel(DeepEvalBaseLLM):
     """Class that implements Google Gemini models for text-based evaluation.
@@ -145,6 +153,7 @@ class GeminiModel(DeepEvalBaseLLM):
         ]
         return self.client.models
 
+    @retry_gemini
     def generate(self, prompt: str, schema: Optional[BaseModel] = None) -> str:
         """Generates text from a prompt.
 
@@ -180,6 +189,7 @@ class GeminiModel(DeepEvalBaseLLM):
             )
             return response.text, 0
 
+    @retry_gemini
     async def a_generate(
         self, prompt: str, schema: Optional[BaseModel] = None
     ) -> str:

deepeval/models/llms/grok_model.py

@@ -1,10 +1,20 @@
+import os
+
 from typing import Optional, Tuple, Union, Dict
 from pydantic import BaseModel
-import os
 
+from deepeval.models.retry_policy import (
+    create_retry_decorator,
+    sdk_retries_for,
+)
 from deepeval.key_handler import ModelKeyValues, KEY_FILE_HANDLER
 from deepeval.models.llms.utils import trim_and_load_json
 from deepeval.models import DeepEvalBaseLLM
+from deepeval.constants import ProviderSlug as PS
+
+
+# consistent retry rules
+retry_grok = create_retry_decorator(PS.GROK)
 
 
 structured_outputs_models = [
@@ -81,6 +91,7 @@ class GrokModel(DeepEvalBaseLLM):
     # Other generate functions
     ###############################################
 
+    @retry_grok
     def generate(
         self, prompt: str, schema: Optional[BaseModel] = None
     ) -> Tuple[Union[str, Dict], float]:
@@ -118,6 +129,7 @@ class GrokModel(DeepEvalBaseLLM):
         else:
             return output, cost
 
+    @retry_grok
     async def a_generate(
         self, prompt: str, schema: Optional[BaseModel] = None
     ) -> Tuple[Union[str, Dict], float]:
@@ -178,9 +190,9 @@ class GrokModel(DeepEvalBaseLLM):
             from xai_sdk import Client, AsyncClient
 
             if not async_mode:
-                return Client(api_key=self.api_key, **self.kwargs)
+                return self._build_client(Client)
             else:
-                return AsyncClient(api_key=self.api_key, **self.kwargs)
+                return self._build_client(AsyncClient)
         except ImportError:
             raise ImportError(
                 "xai_sdk is required to use GrokModel. Please install it with: pip install xai-sdk"
@@ -188,3 +200,38 @@ class GrokModel(DeepEvalBaseLLM):
 
     def get_model_name(self):
         return f"{self.model_name}"
+
+    def _client_kwargs(self) -> Dict:
+        """
+        If Tenacity is managing retries, disable gRPC channel retries to avoid double retry.
+        If the user opts into SDK retries for 'grok' via DEEPEVAL_SDK_RETRY_PROVIDERS,
+        leave channel options as is
+        """
+        kwargs = dict(self.kwargs or {})
+        opts = list(kwargs.get("channel_options", []))
+        if not sdk_retries_for(PS.GROK):
+            # remove any explicit enable flag, then disable retries
+            opts = [
+                option
+                for option in opts
+                if not (
+                    isinstance(option, (tuple, list))
+                    and option
+                    and option[0] == "grpc.enable_retries"
+                )
+            ]
+            opts.append(("grpc.enable_retries", 0))
+        if opts:
+            kwargs["channel_options"] = opts
+        return kwargs
+
+    def _build_client(self, cls):
+        kw = dict(api_key=self.api_key, **self._client_kwargs())
+        try:
+            return cls(**kw)
+        except TypeError as e:
+            # fallback: older SDK version might not accept channel_options
+            if "channel_options" in str(e):
+                kw.pop("channel_options", None)
+                return cls(**kw)
+            raise

deepeval/models/llms/kimi_model.py

@@ -2,11 +2,18 @@ from typing import Optional, Tuple, Union, Dict
 from openai import OpenAI, AsyncOpenAI
 from pydantic import BaseModel
 
+from deepeval.models.retry_policy import (
+    create_retry_decorator,
+    sdk_retries_for,
+)
 from deepeval.key_handler import ModelKeyValues, KEY_FILE_HANDLER
 from deepeval.models.llms.utils import trim_and_load_json
 from deepeval.models import DeepEvalBaseLLM
+from deepeval.constants import ProviderSlug as PS
 
 
+retry_kimi = create_retry_decorator(PS.KIMI)
+
 json_mode_models = [
     "kimi-thinking-preview",
     "kimi-k2-0711-preview",
@@ -100,6 +107,7 @@ class KimiModel(DeepEvalBaseLLM):
     # Other generate functions
     ###############################################
 
+    @retry_kimi
     def generate(
         self, prompt: str, schema: Optional[BaseModel] = None
     ) -> Tuple[Union[str, Dict], float]:
@@ -137,6 +145,7 @@ class KimiModel(DeepEvalBaseLLM):
         else:
             return output, cost
 
+    @retry_kimi
     async def a_generate(
         self, prompt: str, schema: Optional[BaseModel] = None
     ) -> Tuple[Union[str, Dict], float]:
@@ -194,13 +203,34 @@ class KimiModel(DeepEvalBaseLLM):
 
     def load_model(self, async_mode: bool = False):
         if not async_mode:
-            return OpenAI(
-                api_key=self.api_key, base_url=self.base_url, **self.kwargs
-            )
-        else:
-            return AsyncOpenAI(
-                api_key=self.api_key, base_url=self.base_url, **self.kwargs
-            )
+            return self._build_client(OpenAI)
+        return self._build_client(AsyncOpenAI)
+
+    def _client_kwargs(self) -> Dict:
+        """
+        If Tenacity is managing retries, force OpenAI SDK retries off to avoid double retries.
+        If the user opts into SDK retries for 'kimi' via DEEPEVAL_SDK_RETRY_PROVIDERS,
+        leave their retry settings as is.
+        """
+        kwargs = dict(self.kwargs or {})
+        if not sdk_retries_for(PS.KIMI):
+            kwargs["max_retries"] = 0
+        return kwargs
+
+    def _build_client(self, cls):
+        kw = dict(
+            api_key=self.api_key,
+            base_url=self.base_url,
+            **self._client_kwargs(),
+        )
+        try:
+            return cls(**kw)
+        except TypeError as e:
+            # older OpenAI SDKs may not accept max_retries, in that case remove and retry once
+            if "max_retries" in str(e):
+                kw.pop("max_retries", None)
+                return cls(**kw)
+            raise
 
     def get_model_name(self):
         return f"{self.model_name}"

deepeval/models/llms/local_model.py

@@ -1,12 +1,20 @@
 from typing import Optional, Tuple, Union, Dict
 from pydantic import BaseModel
-
 from openai import OpenAI, AsyncOpenAI
 from openai.types.chat import ChatCompletion
 
+from deepeval.models.retry_policy import (
+    create_retry_decorator,
+    sdk_retries_for,
+)
 from deepeval.models.llms.utils import trim_and_load_json
 from deepeval.models import DeepEvalBaseLLM
 from deepeval.key_handler import ModelKeyValues, KEY_FILE_HANDLER
+from deepeval.constants import ProviderSlug as PS
+
+
+# consistent retry rules
+retry_local = create_retry_decorator(PS.LOCAL)
 
 
 class LocalModel(DeepEvalBaseLLM):
@@ -43,6 +51,7 @@ class LocalModel(DeepEvalBaseLLM):
     # Other generate functions
     ###############################################
 
+    @retry_local
     def generate(
         self, prompt: str, schema: Optional[BaseModel] = None
     ) -> Tuple[Union[str, Dict], float]:
@@ -61,6 +70,7 @@ class LocalModel(DeepEvalBaseLLM):
         else:
             return res_content, 0.0
 
+    @retry_local
     async def a_generate(
         self, prompt: str, schema: Optional[BaseModel] = None
     ) -> Tuple[Union[str, Dict], float]:
@@ -91,14 +101,30 @@ class LocalModel(DeepEvalBaseLLM):
 
     def load_model(self, async_mode: bool = False):
         if not async_mode:
-            return OpenAI(
-                api_key=self.local_model_api_key,
-                base_url=self.base_url,
-                **self.kwargs,
-            )
-        else:
-            return AsyncOpenAI(
-                api_key=self.local_model_api_key,
-                base_url=self.base_url,
-                **self.kwargs,
-            )
+            return self._build_client(OpenAI)
+        return self._build_client(AsyncOpenAI)
+
+    def _client_kwargs(self) -> Dict:
+        """
+        If Tenacity manages retries, turn off OpenAI SDK retries to avoid double retrying.
+        If users opt into SDK retries via DEEPEVAL_SDK_RETRY_PROVIDERS=local, leave them enabled.
+        """
+        kwargs = dict(self.kwargs or {})
+        if not sdk_retries_for(PS.LOCAL):
+            kwargs["max_retries"] = 0
+        return kwargs
+
+    def _build_client(self, cls):
+        kw = dict(
+            api_key=self.local_model_api_key,
+            base_url=self.base_url,
+            **self._client_kwargs(),
+        )
+        try:
+            return cls(**kw)
+        except TypeError as e:
+            # Older OpenAI SDKs may not accept max_retries; drop and retry once.
+            if "max_retries" in str(e):
+                kw.pop("max_retries", None)
+                return cls(**kw)
+            raise

deepeval/models/llms/ollama_model.py

@@ -2,8 +2,16 @@ from ollama import Client, AsyncClient, ChatResponse
 from typing import Optional, Tuple, Union, Dict
 from pydantic import BaseModel
 
+from deepeval.models.retry_policy import (
+    create_retry_decorator,
+)
+
 from deepeval.models import DeepEvalBaseLLM
 from deepeval.key_handler import ModelKeyValues, KEY_FILE_HANDLER
+from deepeval.constants import ProviderSlug as PS
+
+
+retry_ollama = create_retry_decorator(PS.OLLAMA)
 
 
 class OllamaModel(DeepEvalBaseLLM):
@@ -34,6 +42,7 @@ class OllamaModel(DeepEvalBaseLLM):
     # Other generate functions
     ###############################################
 
+    @retry_ollama
     def generate(
         self, prompt: str, schema: Optional[BaseModel] = None
     ) -> Tuple[Union[str, Dict], float]:
@@ -56,6 +65,7 @@ class OllamaModel(DeepEvalBaseLLM):
             0,
         )
 
+    @retry_ollama
     async def a_generate(
         self, prompt: str, schema: Optional[BaseModel] = None
     ) -> Tuple[str, float]:
@@ -84,9 +94,11 @@ class OllamaModel(DeepEvalBaseLLM):
 
     def load_model(self, async_mode: bool = False):
         if not async_mode:
-            return Client(host=self.base_url, **self.kwargs)
-        else:
-            return AsyncClient(host=self.base_url, **self.kwargs)
+            return self._build_client(Client)
+        return self._build_client(AsyncClient)
+
+    def _build_client(self, cls):
+        return cls(host=self.base_url, **self.kwargs)
 
     def get_model_name(self):
         return f"{self.model_name} (Ollama)"