fix-openai-toolcall-after-thinking #20333

llama-index-integrations/llms/llama-index-llms-openai/llama_index/llms/openai/responses.py

@@ -536,6 +536,7 @@ def _chat(self, messages: Sequence[ChatMessage], **kwargs: Any) -> ChatResponse:
             messages,
             model=self.model,
             is_responses_api=True,
+            store=kwargs_dict.get("store"),
         )
 
         response: Response = self._client.responses.create(

llama-index-integrations/llms/llama-index-llms-openai/llama_index/llms/openai/utils.py

@@ -352,6 +352,7 @@ def to_openai_message_dict(
     message: ChatMessage,
     drop_none: bool = False,
     model: Optional[str] = None,
+    store: bool = False,
 ) -> ChatCompletionMessageParam:
     """Convert a ChatMessage to an OpenAI message dict."""
     content = []
@@ -533,6 +534,7 @@ def to_openai_responses_message_dict(
     message: ChatMessage,
     drop_none: bool = False,
     model: Optional[str] = None,
+    store: bool = False,
 ) -> Union[str, Dict[str, Any], List[Dict[str, Any]]]:
     """Convert a ChatMessage to an OpenAI message dict."""
     content = []
@@ -581,17 +583,20 @@ def to_openai_responses_message_dict(
                         "detail": block.detail or "auto",
                     }
                 )
+
+        # Omit reasoning items when store is set to False
         elif isinstance(block, ThinkingBlock):
-            if block.content and "id" in block.additional_information:
-                reasoning.append(
-                    {
-                        "type": "reasoning",
-                        "id": block.additional_information["id"],
-                        "summary": [
-                            {"type": "summary_text", "text": block.content or ""}
-                        ],
-                    }
-                )
+            if store:
+                if block.content and "id" in block.additional_information:
+                    reasoning.append(
+                        {
+                            "type": "reasoning",
+                            "id": block.additional_information["id"],
+                            "summary": [
+                                {"type": "summary_text", "text": block.content or ""}
+                            ],
+                        }
+                    )
         elif isinstance(block, ToolCallBlock):
             tool_calls.extend(
                 [
@@ -698,6 +703,7 @@ def to_openai_message_dicts(
     drop_none: bool = False,
     model: Optional[str] = None,
     is_responses_api: bool = False,
+    store: bool = False,
 ) -> Union[List[ChatCompletionMessageParam], str]:
     """Convert generic messages to OpenAI message dicts."""
     if is_responses_api:
@@ -707,6 +713,7 @@ def to_openai_message_dicts(
                 message,
                 drop_none=drop_none,
                 model="o3-mini",  # hardcode to ensure developer messages are used
+                store=store,
             )
             if isinstance(message_dicts, list):
                 final_message_dicts.extend(message_dicts)
@@ -730,6 +737,7 @@ def to_openai_message_dicts(
                 message,
                 drop_none=drop_none,
                 model=model,
+                store=store,
             )
             for message in messages
         ]

llama-index-integrations/llms/llama-index-llms-openai/pyproject.toml

@@ -27,7 +27,7 @@ dev = [
 
 [project]
 name = "llama-index-llms-openai"
-version = "0.6.23"
+version = "0.6.24"
 description = "llama-index llms openai integration"
 authors = [{name = "llama-index"}]
 requires-python = ">=3.9,<4.0"

llama-index-integrations/llms/llama-index-llms-openai/tests/test_openai_responses.py

@@ -800,6 +800,73 @@ def test_messages_to_openai_responses_messages():
         ),
     ]
     openai_messages = to_openai_message_dicts(messages, is_responses_api=True)
+    assert len(openai_messages) == 7
+    assert openai_messages[0]["role"] == "developer"
+    assert openai_messages[0]["content"] == "You are a helpful assistant."
+    assert openai_messages[1]["role"] == "user"
+    assert openai_messages[1]["content"] == "What is the capital of France?"
+    assert openai_messages[2] == {
+        "type": "function_call",
+        "arguments": "{'state': 'France'}",
+        "call_id": "1",
+        "name": "get_capital_city_by_state",
+    }
+    assert openai_messages[3]["role"] == "assistant"
+    assert openai_messages[3]["content"] == "Paris"
+    assert openai_messages[4]["role"] == "user"
+    assert openai_messages[4]["content"] == "What is the capital of Germany?"
+    assert openai_messages[5] == {
+        "type": "function_call",
+        "arguments": "{'state': 'Germany'}",
+        "call_id": "2",
+        "name": "get_capital_city_by_state",
+    }
+    assert openai_messages[6]["role"] == "assistant"
+    assert len(openai_messages[6]["content"]) == 1
+    assert openai_messages[6]["content"][0]["text"] == messages[6].blocks[1].text
+
+
+def test_messages_to_openai_responses_messages_with_store():
+    messages = [
+        ChatMessage(role=MessageRole.SYSTEM, content="You are a helpful assistant."),
+        ChatMessage(role=MessageRole.USER, content="What is the capital of France?"),
+        ChatMessage(
+            role=MessageRole.ASSISTANT,
+            blocks=[
+                ToolCallBlock(
+                    tool_call_id="1",
+                    tool_name="get_capital_city_by_state",
+                    tool_kwargs="{'state': 'France'}",
+                )
+            ],
+        ),
+        ChatMessage(role=MessageRole.ASSISTANT, content="Paris"),
+        ChatMessage(role=MessageRole.USER, content="What is the capital of Germany?"),
+        ChatMessage(
+            role=MessageRole.ASSISTANT,
+            blocks=[
+                ToolCallBlock(
+                    tool_call_id="2",
+                    tool_name="get_capital_city_by_state",
+                    tool_kwargs="{'state': 'Germany'}",
+                )
+            ],
+        ),
+        ChatMessage(
+            role=MessageRole.ASSISTANT,
+            blocks=[
+                ThinkingBlock(
+                    content="The user is asking a simple question related to the capital of Germany, I should answer it concisely",
+                    additional_information={"id": "123456789"},
+                ),
+                TextBlock(text="Berlin"),
+            ],
+        ),
+    ]
+
+    openai_messages = to_openai_message_dicts(
+        messages, is_responses_api=True, store=True
+    )
     assert len(openai_messages) == 8
     assert openai_messages[0]["role"] == "developer"
     assert openai_messages[0]["content"] == "You are a helpful assistant."