Refactor Azure OpenAI response_format validation based on PR feedback

Co-Authored-By: Joe Moura <joao@crewai.com>

src/crewai/llm.py

@@ -127,6 +127,9 @@ def suppress_warnings():
 
 
 class LLM:
+    # Azure OpenAI models that support JSON mode
+    AZURE_JSON_SUPPORTED_MODELS = ["gpt-35-turbo", "gpt-4-turbo", "gpt-4o"]
+    
     def __init__(
         self,
         model: str,
@@ -446,17 +449,42 @@ class LLM:
         if "/" in self.model:
             return self.model.split("/")[0]
         return "openai"
+        
+    def _is_azure_json_supported_model(self) -> bool:
+        """
+        Check if the current model is an Azure OpenAI model that supports JSON mode.
+        
+        Returns:
+            bool: True if the model is an Azure OpenAI model that supports JSON mode, False otherwise.
+        """
+        return any(prefix in self.model for prefix in self.AZURE_JSON_SUPPORTED_MODELS)
 
     def _validate_call_params(self) -> None:
         """
         Validate parameters before making a call. Currently this only checks if
         a response_format is provided and whether the model supports it.
+        
+        Special handling for Azure OpenAI models that support JSON mode:
+        - gpt-35-turbo
+        - gpt-4-turbo
+        - gpt-4o
+        
         The custom_llm_provider is dynamically determined from the model:
           - E.g., "openrouter/deepseek/deepseek-chat" yields "openrouter"
           - "gemini/gemini-1.5-pro" yields "gemini"
           - If no slash is present, "openai" is assumed.
+          
+        Raises:
+            ValueError: If response_format is used with unsupported models
         """
         provider = self._get_custom_llm_provider()
+        
+        # Special case for Azure OpenAI models that support JSON mode
+        if (provider == "azure" and 
+            self.response_format is not None and 
+            self._is_azure_json_supported_model()):
+            return  # Skip validation for Azure OpenAI models that support JSON mode
+            
         if self.response_format is not None and not supports_response_schema(
             model=self.model,
             custom_llm_provider=provider,

tests/llm_test.py

@@ -253,6 +253,45 @@ def test_validate_call_params_no_response_format():
     llm._validate_call_params()
 
 
+class TestAzureOpenAIResponseFormat:
+    @pytest.mark.parametrize("model_name", [
+        "azure/gpt-4o",
+        "azure/gpt-35-turbo",
+        "azure/gpt-4-turbo"
+    ])
+    def test_supported_models(self, model_name):
+        """Test that Azure OpenAI models that should support JSON mode can use response_format."""
+        llm = LLM(model=model_name, response_format={"type": "json_object"})
+        # Should not raise any error
+        llm._validate_call_params()
+    
+    def test_unsupported_model(self):
+        """Test that Azure OpenAI models that don't support JSON mode cannot use response_format."""
+        with patch("crewai.llm.supports_response_schema", return_value=False):
+            llm = LLM(model="azure/text-davinci-003", response_format={"type": "json_object"})
+            with pytest.raises(ValueError) as excinfo:
+                llm._validate_call_params()
+            assert "does not support response_format" in str(excinfo.value)
+    
+    def test_validate_call_params_azure_invalid_response_format(self):
+        """Test that Azure OpenAI models validate the response_format type."""
+        # This test is a placeholder as the current implementation doesn't validate the response_format type
+        # If validation is added in the future, this test should be updated
+        llm = LLM(model="azure/gpt-4-turbo", response_format={"type": "json_object"})
+        # Should not raise any error
+        llm._validate_call_params()
+    
+    def test_validate_call_params_azure_none_provider(self):
+        """Test that non-Azure models with Azure model names don't skip validation."""
+        # This test verifies that the provider check is working correctly
+        with patch("crewai.llm.LLM._get_custom_llm_provider", return_value="openai"):
+            with patch("crewai.llm.supports_response_schema", return_value=False):
+                llm = LLM(model="gpt-4-turbo", response_format={"type": "json_object"})
+                with pytest.raises(ValueError) as excinfo:
+                    llm._validate_call_params()
+                assert "does not support response_format" in str(excinfo.value)
+
+
 @pytest.mark.vcr(filter_headers=["authorization"])
 def test_o3_mini_reasoning_effort_high():
     llm = LLM(