Vapi is the Voice AI platform for developers.

Vapi AI

having issues with my vapi agents (using gpt4-o), saying complete nonsense and not adhering to a prompt that worked earlier. It also repeats phrases on nearly every call which is very strange. any known issue that causes this or any way to fix it? it seriously degrades the performance when the assistant goes off on a complete tangent or switches roles and answers it's own questions mid call lol

<@1150445061390221323>
I'm sorry to hear that you're experiencing issues with your VAPI agents. It sounds like there might be a problem with the prompt configuration or a potential bug affecting the behavior of the agents.

To address this issue, I would recommend the following steps:

1. **Review Prompt Configuration**: Ensure that the prompts you are using are correctly formatted and have not been inadvertently altered since they last worked correctly.

2. **Check for Updates**: Sometimes, issues can arise from outdated software. Make sure that your VAPI integration is up to date with the latest version.

3. **Inspect Logs**: Look at the logs for the calls where the agents are not performing as expected. This can provide insights into what might be going wrong.

4. **Test with Simplified Prompts**: Try using simplified prompts to see if the issue persists. This can help isolate whether the problem is with the complexity of the prompts or something else.

5. **Contact Support**: If the above steps do not resolve the issue, it may be necessary to contact support for further assistance. They can provide more detailed troubleshooting and potentially identify if there's a known issue.

If you'd like, I can assist you in closing this support ticket once the issue is resolved, or if you need to escalate this to a higher level of support, please let me know.

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

i removed KB things and it seems to be doing better at least

Isn't Agents repeating phrases means something wrong with prompt?

Yeah. But, not in this case because we have a bug in KB due to which two streams are getting merged which people see as a hallucination issue.

This is interesting media streams are getting merged, teleporting to different calls.

To elaborate further, we subscribe to two concurrent streams. The first stream is for the knowledge base part, and the second stream is for LLM request completion. These streams occasionally merge, which causes this issue to occur.

Oh I misunderstood, by stream i though twilio/vongage media streams.