Returning function call's responses in `raw_response_event`

[levulinh]

I am currently trying to build a chat web app for my agent, which is hosted using fastapi. Since the UI library I'm using accepts the OpenAI Responses API format, I return exactly what is yielded in raw_response_event. However, I found that there's no way to retrieve the response from my custom function calls, so I can't build a complete message history with function call results. I understand that I can explicitly get the function's result from run_item_stream_event, but is there a way to not introducing a new type of event in the stream?
Does anyone have a solution for this? And how are you building a web UI for the agents?
Any suggestions are much appreciated. Thanks.

[neuronwave]

Not sure if this is exactly what you are after, but if you have the function call return an object (dataclass for eg) and set the output_type of the agent making the call to the dataclass Class type then the results returned from the agent are not reinterpreted by the LLM.

So in a weather_agent you could have the following code to set up the function call

class Weather(BaseModel):
  city: str
  temperature_range: str
  conditions: str

@function_tool
def get_weather(ctx: RunContextWrapper, city: str | None) -> Weather:
  print("[debug] get_weather called")
  print(city)
  return Weather(city=city, temperature_range="14-20C", conditions="Sunny with wind")

and then

when you instantiate the weather_agent you can pass in the output_type eg.

class WeatherAgent(Agent):

  def __init__(self, raw=False):
    super().__init__(
      name="Weather agent",
      instructions=(
        "You are an agent who can get the weather"
      ),
      handoff_description="A weather predictor that can provide accurate forecasts",
      tools=[get_weather],
      output_type = Weather

[levulinh]

Sorry for my question might be confusing. What I want to ask is that when I use this to restream the raw output from the LLM so the stream can be read by openai.responses():

 async for event in stream.stream_events():
        if event.type == "raw_response_event":
            event_data = event.data
            event_type = event_data.type
            yield "event: " + event_type + "\n"
            yield "data: " + event_data.model_dump_json() + "\n\n"

There isn't any official event type that indicates the finish of a function call (https://platform.openai.com/docs/api-reference/responses-streaming/response/output_text). The closest thing may be response.output_item.done, but it also doesn't include the results from the function call. So, I wonder if there's a way to provide the results of the functions within just raw_response_event event type.

[abbas-khaku]

@levulinh Would a use case for this be to retrieve the id of the previous conversation?

[levulinh]

@abbas-khaku I suppose handling the history on the server side would work in this case. However, my ui library doesn't seem to support including the message ids in the request. I think I should rephrase my question to "Is there any good React UI library that works with Agents SDK?".

[github-actions]

This issue is stale because it has been open for 7 days with no activity.

[github-actions]

This issue was closed because it has been inactive for 3 days since being marked as stale.