Python: Send the NVIDIA nvext setting in extra_body

python/semantic_kernel/connectors/ai/nvidia/prompt_execution_settings/nvidia_prompt_execution_settings.py

@@ -65,9 +65,17 @@ class NvidiaChatPromptExecutionSettings(NvidiaPromptExecutionSettings):
     nvext: dict[str, Any] | None = None
 
     def prepare_settings_dict(self, **kwargs) -> dict[str, Any]:
-        """Override only for embeddings to exclude input_type and truncate."""
-        return self.model_dump(
-            exclude={"service_id", "extension_data", "structured_json_response", "response_format"},
+        """Prepare the settings dict, moving nvext into extra_body.
+
+        The OpenAI client only sends nvext as part of extra_body. Keys already in extra_body["nvext"],
+        such as the guided_json set from response_format, take precedence over the same keys in nvext.
+        """
+        settings_dict = self.model_dump(
+            exclude={"service_id", "extension_data", "structured_json_response", "response_format", "nvext"},
             exclude_none=True,
             by_alias=True,
         )
+        if self.nvext:
+            extra_body = settings_dict.setdefault("extra_body", {})
+            extra_body["nvext"] = self.nvext | extra_body.get("nvext", {})
+        return settings_dict

python/semantic_kernel/connectors/ai/nvidia/services/nvidia_handler.py

@@ -66,13 +66,6 @@ async def _send_chat_completion_request(
         """Send a request to the NVIDIA chat completion endpoint."""
         try:
             settings_dict = settings.prepare_settings_dict()
-
-            # Handle structured output if nvext is present in extra_body
-            if settings.extra_body and "nvext" in settings.extra_body:
-                if "extra_body" not in settings_dict:
-                    settings_dict["extra_body"] = {}
-                settings_dict["extra_body"]["nvext"] = settings.extra_body["nvext"]
-
             response = await self.client.chat.completions.create(**settings_dict)
             self.store_usage(response)
             return response

python/tests/unit/connectors/ai/nvidia/services/test_nvidia_chat_completion.py

@@ -136,6 +136,46 @@ class TestModel(BaseModel):
         assert "nvext" in call_args["extra_body"]
         assert "guided_json" in call_args["extra_body"]["nvext"]
 
+    @pytest.mark.asyncio
+    @patch.object(AsyncCompletions, "create", new_callable=AsyncMock)
+    async def test_nvext_is_sent_in_extra_body(self, mock_create, nvidia_unit_test_env):
+        """Test that the nvext setting is sent in extra_body, not as a create() argument."""
+        mock_create.return_value = _create_mock_chat_completion()
+
+        service = NvidiaChatCompletion()
+        chat_history = ChatHistory()
+        chat_history.add_user_message("Hello")
+        settings = NvidiaChatPromptExecutionSettings(nvext={"guided_choice": ["yes", "no"]})
+
+        await service.get_chat_message_contents(chat_history, settings)
+
+        call_args = mock_create.call_args[1]
+        assert "nvext" not in call_args
+        assert call_args["extra_body"]["nvext"] == {"guided_choice": ["yes", "no"]}
+
+    @pytest.mark.asyncio
+    @patch.object(AsyncCompletions, "create", new_callable=AsyncMock)
+    async def test_nvext_is_merged_with_structured_output(self, mock_create, nvidia_unit_test_env):
+        """Test that nvext keys are kept next to the guided_json set from response_format."""
+
+        class TestModel(BaseModel):
+            name: str
+
+        mock_create.return_value = _create_mock_chat_completion('{"name": "test"}')
+
+        service = NvidiaChatCompletion()
+        chat_history = ChatHistory()
+        chat_history.add_user_message("Give me structured data")
+        settings = NvidiaChatPromptExecutionSettings(nvext={"repetition_penalty": 1.1}, response_format=TestModel)
+
+        await service.get_chat_message_contents(chat_history, settings)
+
+        call_args = mock_create.call_args[1]
+        assert call_args["extra_body"]["nvext"] == {
+            "repetition_penalty": 1.1,
+            "guided_json": TestModel.model_json_schema(),
+        }
+
     @pytest.mark.asyncio
     @patch.object(AsyncCompletions, "create", new_callable=AsyncMock)
     async def test_error_handling(self, mock_create, nvidia_unit_test_env):