Fix output_json with custom OpenAI APIs by using PARALLEL_TOOLS mode

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

src/crewai/utilities/internal_instructor.py

@@ -29,7 +29,14 @@ class InternalInstructor:
             import instructor
             from litellm import completion
 
-            self._client = instructor.from_litellm(completion)
+            is_custom_openai = getattr(self.llm, 'model', '').startswith('custom_openai/')
+
+            mode = instructor.Mode.PARALLEL_TOOLS if is_custom_openai else instructor.Mode.TOOLS
+
+            self._client = instructor.from_litellm(
+                completion,
+                mode=mode,
+            )
 
     def to_json(self):
         model = self.to_pydantic()
@@ -40,4 +47,8 @@ class InternalInstructor:
         model = self._client.chat.completions.create(
             model=self.llm.model, response_model=self.model, messages=messages
         )
+        
+        if isinstance(model, list) and len(model) > 0:
+            return model[0]  # Return the first model from the list
+            
         return model

tests/utilities/test_internal_instructor.py

@@ -0,0 +1,72 @@
+import unittest
+from unittest.mock import MagicMock, patch
+
+import pytest
+from pydantic import BaseModel
+
+from crewai.utilities.internal_instructor import InternalInstructor
+
+
+class TestOutput(BaseModel):
+    value: str
+
+
+class TestInternalInstructor(unittest.TestCase):
+    @patch("instructor.from_litellm")
+    def test_tools_mode_for_regular_models(self, mock_from_litellm):
+        mock_llm = MagicMock()
+        mock_llm.model = "gpt-4o"
+        mock_instructor = MagicMock()
+        mock_from_litellm.return_value = mock_instructor
+        
+        instructor = InternalInstructor(
+            content="Test content",
+            model=TestOutput,
+            llm=mock_llm
+        )
+        
+        import instructor
+        mock_from_litellm.assert_called_once_with(
+            unittest.mock.ANY,
+            mode=instructor.Mode.TOOLS
+        )
+    
+    @patch("instructor.from_litellm")
+    def test_parallel_tools_mode_for_custom_openai(self, mock_from_litellm):
+        mock_llm = MagicMock()
+        mock_llm.model = "custom_openai/some-model"
+        mock_instructor = MagicMock()
+        mock_from_litellm.return_value = mock_instructor
+        
+        instructor = InternalInstructor(
+            content="Test content",
+            model=TestOutput,
+            llm=mock_llm
+        )
+        
+        import instructor
+        mock_from_litellm.assert_called_once_with(
+            unittest.mock.ANY,
+            mode=instructor.Mode.PARALLEL_TOOLS
+        )
+    
+    @patch("instructor.from_litellm")
+    def test_handling_list_response_in_to_pydantic(self, mock_from_litellm):
+        mock_llm = MagicMock()
+        mock_llm.model = "custom_openai/some-model"
+        mock_instructor = MagicMock()
+        mock_chat = MagicMock()
+        mock_instructor.chat.completions.create.return_value = [
+            TestOutput(value="test value")
+        ]
+        mock_from_litellm.return_value = mock_instructor
+        
+        instructor = InternalInstructor(
+            content="Test content",
+            model=TestOutput,
+            llm=mock_llm
+        )
+        result = instructor.to_pydantic()
+        
+        assert isinstance(result, TestOutput)
+        assert result.value == "test value"