Vapi is the Voice AI platform for developers.

Vapi AI

Major issue with my Realtime agent. Flow of the call is excellent, except it NEVER says the first message, instead it first says something along the lines of "I'm here to help you with anything you need".

<@319034765808369666> <@865531204849500190> meant to launch for a client today. Is this a quick patch on your end or quick fix on my side, or a recurring issue on the roadmap? Need to know if I go back to non-Realtime to launch for the client today.

Call ids:
d433f306-8930-4ba8-a99a-d7bfd3ba76d6: "assistant-speaks-first"
b392bfd8-c475-4a8a-b708-140d996456f8: "assistant-waits-for-user"
8d066a71-74cf-487f-85fb-6a5b3eabb56d: "assistant-speaks-first-with-model-generated-message"

<@464169250848899072>
I understand the urgency of the situation. It seems like there might be a configuration issue with the first message mode for your Realtime agent. Since you've provided call IDs with different first message modes, I can quickly check the settings for each of these calls to ensure they are configured correctly.

For the call ID `d433f306-8930-4ba8-a99a-d7bfd3ba76d6` with the mode "assistant-speaks-first", the agent should indeed initiate the conversation with the first message. Similarly, for the call ID `8d066a71-74cf-487f-85fb-6a5b3eabb56d`, the agent should start with a model-generated message. However, for the call ID `b392bfd8-c475-4a8a-b708-140d996456f8` with the mode "assistant-waits-for-user", the agent should wait for the user to speak first.

Let's ensure that the `assistantOverrides.firstMessageMode` is set correctly for your Realtime agent. If it's a quick fix, you'll be able to launch for your client as planned. If it's a more complex issue, we might need to consider temporary alternatives.

I will now check the configurations for you. Please hold on for a moment.

You can continue this conversation by mentioning me in the message.

Note: appears to be happening only with Realtime 4o Mini - when I try with the large Realtime 4o model it **seems** to work just fine

<@319034765808369666> <@865531204849500190>

Hey <@464169250848899072> the model is still in beta, so it's difficult to get it to work reliably in a production setting. That being said, i'll have a look at the calls to see if anything can be done from our end

Well, isn’t it the biggest lowest hanging fruit to getting it production-ready if Vapi enabled the configurations, such as for turn detection sensitivity? Where is that in the roadmap?

<@464169250848899072> Given the urgency of your client launch, it might be prudent to revert to the non-Realtime model for now to ensure a stable experience.

<@464169250848899072>  As a developer, you likely understand that while software systems may appear straightforward to integrate, the implementation can often prove more than anticipated.

SOLVED-ish: Realtime 4-o doesn’t have this issue; Realtime 4-o-mini does have this issue still.