fix #20333 when store=False

Author: martinalupini

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,
+            kwargs_dict=kwargs_dict,
         )
 
         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,
+    kwargs: Optional[Dict[str, Any]] = None,
 ) -> 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,
+    kwargs: Optional[Dict[str, Any]] = None,
 ) -> Union[str, Dict[str, Any], List[Dict[str, Any]]]:
     """Convert a ChatMessage to an OpenAI message dict."""
     content = []
@@ -582,10 +584,21 @@ def to_openai_responses_message_dict(
                     }
                 )
 
-        # Omit reasoning items from the conversation history
+        # Omit reasoning items when store is set to False
         elif isinstance(block, ThinkingBlock):
-            continue
-
+            if kwargs is None:
+                continue
+            elif kwargs["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(
                 [
@@ -692,6 +705,7 @@ def to_openai_message_dicts(
     drop_none: bool = False,
     model: Optional[str] = None,
     is_responses_api: bool = False,
+    kwargs: Optional[Dict[str, Any]] = None,
 ) -> Union[List[ChatCompletionMessageParam], str]:
     """Convert generic messages to OpenAI message dicts."""
     if is_responses_api:
@@ -701,6 +715,7 @@ def to_openai_message_dicts(
                 message,
                 drop_none=drop_none,
                 model="o3-mini",  # hardcode to ensure developer messages are used
+                kwargs=kwargs,
             )
             if isinstance(message_dicts, list):
                 final_message_dicts.extend(message_dicts)
@@ -712,9 +727,8 @@ def to_openai_message_dicts(
         # If there is only one message, and it is a user message, return the content string directly
         if (
             len(final_message_dicts) == 1
-            and isinstance(final_message_dicts[0], dict)
-            and final_message_dicts[0].get("role") == "user"
-            and isinstance(final_message_dicts[0].get("content"), str)
+            and final_message_dicts[0]["role"] == "user"
+            and isinstance(final_message_dicts[0]["content"], str)
         ):
             return final_message_dicts[0]["content"]
 
@@ -725,6 +739,7 @@ def to_openai_message_dicts(
                 message,
                 drop_none=drop_none,
                 model=model,
+                kwargs=kwargs,
             )
             for message in messages
         ]

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

@@ -826,6 +826,95 @@ def test_messages_to_openai_responses_messages():
     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"),
+            ],
+        ),
+    ]
+
+    kwargs = {
+        "model": "fake-model",
+        "include": None,
+        "instructions": None,
+        "max_output_tokens": 100,
+        "metadata": {},
+        "previous_response_id": None,
+        "store": True,
+        "temperature": 0.0,
+        "tools": [],
+        "top_p": 1.0,
+        "truncation": None,
+        "user": None,
+    }
+
+    openai_messages = to_openai_message_dicts(
+        messages, is_responses_api=True, kwargs=kwargs
+    )
+    assert len(openai_messages) == 8
+    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]["type"] == "reasoning"
+    assert (
+        openai_messages[6]["id"] == messages[6].blocks[0].additional_information["id"]
+    )
+    assert openai_messages[6]["summary"][0]["text"] == messages[6].blocks[0].content
+
+    assert openai_messages[7]["role"] == "assistant"
+    assert len(openai_messages[7]["content"]) == 1
+    assert openai_messages[7]["content"][0]["text"] == messages[6].blocks[1].text
+
+
 @pytest.fixture()
 def response_output() -> List[ResponseOutputItem]:
     return [

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

@@ -21,7 +21,6 @@
     MessageRole,
     TextBlock,
     ToolCallBlock,
-    ThinkingBlock,
 )
 from llama_index.core.bridge.pydantic import BaseModel
 from llama_index.llms.openai import OpenAI
@@ -502,87 +501,3 @@ def test_gpt_5_chat_model_support() -> None:
     )
 
     assert model_name in CHAT_MODELS, f"{model_name} should be in CHAT_MODELS"
-
-
-def test_to_openai_message_dicts_responses_api_drops_reasoning_before_tool_call() -> (
-    None
-):
-    """
-    Test that 'reasoning items' are not included when converting a ChatMessage to an OpenAI message dict.
-    (they are internal model thinking and should not be included in conversational history)
-    """
-    msg = ChatMessage(
-        role=MessageRole.ASSISTANT,
-        blocks=[
-            ThinkingBlock(
-                content="I will call the tool now.",
-                additional_information={"id": "rs_dummy_reasoning_id"},
-            ),
-            ToolCallBlock(
-                block_type="tool_call",
-                tool_call_id="call_123",
-                tool_name="search_hotels",
-                tool_kwargs='{"location":"Rome","max_price":200}',
-            ),
-        ],
-    )
-
-    out = to_openai_message_dicts([msg], is_responses_api=True)
-
-    assert isinstance(out, list)
-
-    # Must not include any reasoning items in the serialized input
-    assert not any(
-        isinstance(item, dict) and item.get("type") == "reasoning" for item in out
-    )
-
-    # Must still include the tool/function call item
-    assert any(
-        isinstance(item, dict)
-        and item.get("type") == "function_call"
-        and item.get("name") == "search_hotels"
-        and item.get("call_id") == "call_123"
-        for item in out
-    )
-
-
-def test_to_openai_message_dicts_responses_api_drops_reasoning_before_tool_call() -> (
-    None
-):
-    """
-    Test that 'reasoning items' are not included when converting a ChatMessage to an OpenAI message dict.
-    (they are internal model thinking and should not be included in conversational history)
-    """
-    msg = ChatMessage(
-        role=MessageRole.ASSISTANT,
-        blocks=[
-            ThinkingBlock(
-                content="I will call the tool now.",
-                additional_information={"id": "rs_dummy_reasoning_id"},
-            ),
-            ToolCallBlock(
-                block_type="tool_call",
-                tool_call_id="call_123",
-                tool_name="search_hotels",
-                tool_kwargs='{"location":"Rome","max_price":200}',
-            ),
-        ],
-    )
-
-    out = to_openai_message_dicts([msg], is_responses_api=True)
-
-    assert isinstance(out, list)
-
-    # Must not include any reasoning items in the serialized input
-    assert not any(
-        isinstance(item, dict) and item.get("type") == "reasoning" for item in out
-    )
-
-    # Must still include the tool/function call item
-    assert any(
-        isinstance(item, dict)
-        and item.get("type") == "function_call"
-        and item.get("name") == "search_hotels"
-        and item.get("call_id") == "call_123"
-        for item in out
-    )