Assistant started speaking gibberish
# supportt
teamrick
05/05/2024, 9:22 AMHi, today for one of our calls the assistant started saying nonsensical sentences. The prompt has not been changed and so far it's been behaving normally.
Here's the transcript (from Vapi's dashboard)
AI: Hi. This is Sarah from Global University I'm calling regarding your inquiry for our business degree. Am I speaking with Alex?
User: Yeah. This is Alex.
AI: Fantastic. Alex. I'd like to help schedule a call with your dedicated adviser to discuss the next steps. First just need to confirm some details. You have a high school diploma?
User: Yeah. I have a high school diploma.
AI: Great. Great. To Alex here, that that's Alex what? We now need have now you considered do how you you already owe half finance or your plan degree on How to finance your degree,
Call ID: 8f392dda-2322-435c-8c20-dd544eaa537b
Can I check why this happens? It's on GPT 4
v
Vapi Support
05/05/2024, 9:22 AMHi @teamrick, I'm an AI assistant for Vapi. I'll analyze available sources and get back to you shortly.
Vapi Support
05/05/2024, 9:22 AM### Answer
The issue you're experiencing with the assistant speaking in a manner that doesn't make sense could be attributed to a few factors, none of which are directly covered in the information sources provided. However, based on common practices in AI and language models, this could involve:
- **Prompt complexity**: If the prompt or conversation flow is complex, it could cause the model to generate nonsensical responses. Ensuring prompts are clear and structured can help mitigate this issue.
- **Model limitations**: Even with advanced models like GPT-4, unexpected responses can occur due to the inherent unpredictability of AI language models.
Since this issue isn't explicitly addressed in the provided documentation, and considering the limitations of my current knowledge base, I recommend reviewing the setup of your assistant's prompt or the conversation flow to ensure clarity. Additionally, if you have access to model configurations, you may consider adjusting parameters such as temperature to see if it impacts the coherence of the responses.
For detailed assistance or to report this behavior, directly contacting Vapi's support or development team could provide more insights, given they'll have access to logs and more detailed model behavior data.
Vapi Support
05/05/2024, 9:25 AMGot it. Tagging @User for additional help.
s
Sahil
05/05/2024, 12:52 PMCan you lower the temperature and try it once more?
t
teamrick
05/06/2024, 7:50 AMThanks! Could you help check the log to see if the error is completely from the reply of GPT 4, if there was any bug with the prompt passed to it, or to the transcription made by the model?
I've not experienced something like this chatting with GPT 3.5 / 4 so far so it's a bit weird that it was so garbled
teamrick
05/06/2024, 2:30 PM@Sahil thank you 🙂
v
Vapi Support
05/06/2024, 2:30 PMGot it. Tagging @User for additional help.
s
Sahil
05/06/2024, 2:31 PMDid it work or you still have any issue?
t
teamrick
05/06/2024, 2:36 PMit only appeared once randomly out of 20+ calls on the same prompt, but I want to understand why, because the agent speaking gibberish 5% of the time isn't good for our use case
teamrick
05/06/2024, 2:37 PMand do you see this happen with other people?
teamrick
05/06/2024, 2:38 PMI've lowered the temperature but it's hard to know if that solves it or if it's just disappeared for now. Also, do you know if lowering the temp drops it to 0% vs 1% chance of happening?
s
Sam
05/06/2024, 3:03 PMI have the same issue - 11labs voices started being unstable all of the sudden and also i get a lot of errors and sometimes AI agent supports responding!!! and today i got LLM error!
Sam
05/06/2024, 3:04 PMcan someone help me - all my ai agents are down. i haven't changed anything since Friday
2 Views