Are reasoning items of OpenAI returned back when using llama-index agent systems?

[YutackPark]

In this guideline of OpenAI: https://platform.openai.com/docs/guides/reasoning?api-mode=responses

I found this paragraph about using reasoning models with function calling ability

When doing function calling with a reasoning model in the Responses API, we highly recommend you pass back any reasoning items returned with the last function call (in addition to the output of your function). If the model calls multiple functions consecutively, you should pass back all reasoning items, function call items, and function call output items, since the last user message. This allows the model to continue its reasoning process to produce better results in the most token-efficient manner.

As llamaIndex agent came first and then Response API, I'm not sure how it is related and actually works. The paragraph sound like, "If you gonna use reasoning model with function calling, you should use stateful ResponseAPI to get the best performance".

User can not directly ensure or control the behavior as openAI does not support raw reasoning items, but they are somehow encapsulated in ResponseAPI and their id.

[dosubot[bot]]

Yes, LlamaIndex agent systems can return reasoning items from OpenAI when using the newer OpenAI Responses API. The integration is handled through the OpenAIResponses class, which supports reasoning options (like the reasoning parameter for O1 models) and processes reasoning items returned by OpenAI's API. If you enable the track_previous_responses attribute, LlamaIndex will statefully track and pass back previous reasoning items, function call items, and outputs—just as recommended in the OpenAI guidelines for best performance with reasoning models and function calling. This approach aligns with the stateful interaction model described in the OpenAI documentation, even though the raw reasoning items themselves are encapsulated and not directly exposed to the user—they are managed internally by the API and LlamaIndex integration. For more details, see the implementation in OpenAIResponses.

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}   

[YutackPark]

I have found exact "reasoning_item" field from open AI responses which is transferred as additional_kwargs. The logic behind track_previous_responses is about response id, not directly to the additional_kwargs, which confuses me...
I'll update with my experiments later.