Merge pull request #19201 from BerriAI/litellm_fix_vertex_ai_structured_output2

Fix: vertex ai doesn't support structured output

litellm/llms/vertex_ai/vertex_ai_partner_models/anthropic/transformation.py

@@ -69,6 +69,9 @@ class VertexAIAnthropicConfig(AnthropicConfig):
 
         data.pop("model", None)  # vertex anthropic doesn't accept 'model' parameter
         
+        # VertexAI doesn't support output_format parameter, remove it if present
+        data.pop("output_format", None)
+        
         tools = optional_params.get("tools")
         tool_search_used = self.is_tool_search_used(tools)
         auto_betas = self.get_anthropic_beta_list(
@@ -89,6 +92,37 @@ class VertexAIAnthropicConfig(AnthropicConfig):
         
         return data
 
+    def map_openai_params(
+        self,
+        non_default_params: dict,
+        optional_params: dict,
+        model: str,
+        drop_params: bool,
+    ) -> dict:
+        """
+        Override parent method to ensure VertexAI always uses tool-based structured outputs.
+        VertexAI doesn't support the output_format parameter, so we force all models
+        to use the tool-based approach for structured outputs.
+        """
+        # Temporarily override model name to force tool-based approach
+        # This ensures Claude Sonnet 4.5 uses tools instead of output_format
+        original_model = model
+        if "response_format" in non_default_params:
+            model = "claude-3-sonnet-20240229"  # Use a model that will use tool-based approach
+        
+        # Call parent method with potentially modified model name
+        optional_params = super().map_openai_params(
+            non_default_params=non_default_params,
+            optional_params=optional_params,
+            model=model,
+            drop_params=drop_params,
+        )
+        
+        # Restore original model name for any other processing
+        model = original_model
+        
+        return optional_params
+
     def transform_response(
         self,
         model: str,

tests/test_litellm/llms/vertex_ai/vertex_ai_partner_models/anthropic/test_vertex_ai_partner_models_anthropic_transformation.py

@@ -115,3 +115,147 @@ def test_vertex_ai_anthropic_structured_output_header_not_added():
         "Non-Vertex request SHOULD have anthropic-beta header for structured output"
     assert result_non_vertex["anthropic-beta"] == "structured-outputs-2025-11-13", \
         f"Expected 'structured-outputs-2025-11-13', got: {result_non_vertex.get('anthropic-beta')}"
+
+
+def test_vertex_ai_claude_sonnet_4_5_structured_output_fix():
+    """
+    Test fix for issue #18625: Claude Sonnet 4.5 on VertexAI should use tool-based 
+    structured outputs instead of output_format parameter.
+    
+    This test verifies that:
+    1. Claude Sonnet 4.5 uses tool-based structured outputs on VertexAI
+    2. output_format parameter is removed from the final request
+    3. The fix prevents "Extra inputs are not permitted" error
+    """
+    config = VertexAIAnthropicConfig()
+    
+    # Test data matching the issue report
+    response_format = {
+        "type": "json_schema",
+        "json_schema": {
+            "name": "questions",
+            "strict": True,
+            "schema": {
+                "type": "object",
+                "properties": {
+                    "question": {
+                        "type": "string"
+                    },
+                    "response": {
+                        "type": "string"
+                    }
+                },
+                "required": ["question", "response"],
+                "additionalProperties": False
+            }
+        }
+    }
+    
+    messages = [
+        {"role": "user", "content": "Generate a question and answer about AI."}
+    ]
+    
+    # Test parameters that would trigger the issue
+    non_default_params = {
+        "response_format": response_format,
+        "max_tokens": 1000,
+    }
+    
+    # Test 1: Verify map_openai_params forces tool-based approach for Claude Sonnet 4.5
+    optional_params = {}
+    result_params = config.map_openai_params(
+        non_default_params=non_default_params,
+        optional_params=optional_params,
+        model="claude-3-5-sonnet-20241022",  # Claude Sonnet 4.5 model
+        drop_params=False,
+    )
+    
+    # Should have tools and tool_choice (tool-based approach)
+    assert "tools" in result_params, "Tools should be present for structured output"
+    assert "tool_choice" in result_params, "Tool choice should be present for structured output"
+    assert "json_mode" in result_params, "JSON mode should be enabled"
+    
+    # Verify the tool is the response format tool
+    tools = result_params["tools"]
+    assert len(tools) == 1, "Should have exactly one tool for response format"
+    assert tools[0]["name"] == "json_tool_call", "Tool should be named json_tool_call"
+    
+    # Test 2: Verify transform_request removes output_format parameter
+    # Simulate what would happen if parent class added output_format
+    test_data = {
+        "model": "claude-3-5-sonnet-20241022",
+        "messages": messages,
+        "max_tokens": 1000,
+        "tools": tools,
+        "tool_choice": result_params["tool_choice"],
+        "output_format": {  # This would be added by parent class for Sonnet 4.5
+            "type": "json_schema",
+            "schema": response_format["json_schema"]["schema"]
+        }
+    }
+    
+    # Mock the parent transform_request to return data with output_format
+    original_transform = config.__class__.__bases__[0].transform_request
+    
+    def mock_transform_request(self, model, messages, optional_params, litellm_params, headers):
+        # Return test data that includes output_format
+        return test_data.copy()
+    
+    # Temporarily replace parent method
+    config.__class__.__bases__[0].transform_request = mock_transform_request
+    
+    try:
+        final_data = config.transform_request(
+            model="claude-3-5-sonnet-20241022",
+            messages=messages,
+            optional_params=result_params,
+            litellm_params={},
+            headers={},
+        )
+        
+        # Verify that output_format was removed (fixes the "Extra inputs are not permitted" error)
+        assert "output_format" not in final_data, "output_format should be removed for VertexAI"
+        assert "model" not in final_data, "model should be removed for VertexAI"
+        assert "tools" in final_data, "tools should still be present"
+        assert "tool_choice" in final_data, "tool_choice should still be present"
+        
+    finally:
+        # Restore original method
+        config.__class__.__bases__[0].transform_request = original_transform
+
+
+def test_vertex_ai_anthropic_other_models_still_use_tools():
+    """
+    Test that other Anthropic models (non-Sonnet 4.5) on VertexAI also use tool-based 
+    structured outputs, ensuring consistency across all models.
+    """
+    config = VertexAIAnthropicConfig()
+    
+    response_format = {
+        "type": "json_schema",
+        "json_schema": {
+            "name": "test_schema",
+            "schema": {
+                "type": "object",
+                "properties": {
+                    "result": {"type": "string"}
+                }
+            }
+        }
+    }
+    
+    # Test with Claude 3 Sonnet (not 4.5)
+    non_default_params = {"response_format": response_format}
+    optional_params = {}
+    
+    result_params = config.map_openai_params(
+        non_default_params=non_default_params,
+        optional_params=optional_params,
+        model="claude-3-sonnet-20240229",
+        drop_params=False,
+    )
+    
+    # Should still use tool-based approach
+    assert "tools" in result_params, "Claude 3 Sonnet should also use tool-based structured output"
+    assert "tool_choice" in result_params, "Tool choice should be present"
+    assert "json_mode" in result_params, "JSON mode should be enabled"