Has anyone had problems with gpt-realtime not responding with the tool result?

API
,
1

Hi!

We’re using gpt-realtime with a payload like

            audio: {
              input: {
                turn_detection: {
                  type: "server_vad",
                  create_response: true,
                  interrupt_response: true,
                  silence_duration_ms: 700,
                  idle_timeout_ms: 30000
                },
              },
              output: { voice: VOICE }
            },

and, well it’s hard to explain, but when I ask it to run a tool, it does that but it doesn’t give me the response until I say smth like: “What was the result?” The prompt has specific instructions telling to communicate during the whole lifetime of the tool call, but it feels like it’s waiting for me to say something before it communicates back, even though I’m waiting for it to finish what it’s doing.

I’m guessing I’m doing something wrong, but I can’t figure out what. Has anyone run into a similar situation? Any ideas on what to try?

2

does your tool response tell the model to communicate something to the user?

{
            "type": "conversation.item.create",
            "item": {
                "type": "function_call_output",
                "output": "tell the user about "+str(result),
                "call_id": call_id
            }
        }
3

Will try that, but it feels a bit counterintuitive, no?