This is an issue reported by others. Several times now, during calls, the AI ends up gettings its words and sentences all mixed up resulting in nonsense. I can see there's some meaning behind it all, but the words are in the wrong order. Example from transcript: "AI: Sure, Sam. I get that. That's a fair point. But think about, Sam, it this way. If there were a culture in this scenario of constant, the communication and everyone looking out worker was alone. For each so direct communication with other maybe someone a team else could have noticed the potential danger and called out might not a warning. Better have been a factor. But what communication might about before the incident involve more situational? Maybe awareness across there, could the whole team have been a briefing what or some do you form of think. Could communication that fostering a might have stronger culture highlighted the teamwork risks of and communication? Working near help and running engines, other areas as well. Call ID: c623c7f0-88a8-4543-9b72-68955e3a8039 Call time: 03:05 - 03:46

Hi @Jinani, I'm an AI assistant for Vapi. I'll analyze available sources and get back to you shortly.

Damn feels like I’m having a stroke reading that lol models are weird;

GPT4o. I don’t think it’s a model issue but something to do with the way it’s streamed and processed in Vapi. But yeah, it’s happened frequently enough to be an issue.

Got it. Tagging @User for additional help.

Can you check the LLM Output in the call logs section and tell me whether the words are getting mixed up between the LLM Output and Voice Input parts? If not, then it is not a streaming issue. https://cdn.discordapp.com/attachments/1244745971678838895/1244951237242130492/image.png?ex=6656fab1&is=6655a931&hm=ca90f757076b73f261d00cf6a1615f82b088c71771827f3a362943679d2d2b6a&

Seems like the model output is the same as the voice input, if I'm reading the log correctly? https://cdn.discordapp.com/attachments/1244745971678838895/1245344429523603537/Screenshot_2024-05-29_at_14.52.34.png?ex=665868e1&is=66571761&hm=a77113691ba137df8660c52febd4075342239e35b59de4c976833b56c2ee7876&

Yup. It is same just a little formatting is done before sending to voice input.

So is it an issue with GPT4o? Hopefully something that can be fixed.

Try GPT-4 Turbo

Hi @Sahil and @JM. I think I might be experiencing the same behavior. @JM, do you have RAG data for this assistant? It seems like I can only get it to happen with gpt-4o and a document for rag. AI: Absolutely. Doctor Graso absolutely has been. Doctor treating patients grace 4 has over been treating 17 patients years for over. You're in 17 great years hands, and has heaps of experience. You'll be in good hands. User: Oh, no. You just you just were talking crazy right there. AI: Uh, my bad, Tyler. Got a bit carried away there. Doctor Wang is really good though. Anything else I can help you with? * I love her recovery 🙂 Call Id: 942c4501-5aea-48db-8f27-bbc368816299 Transcript: 16:21:35:869 Audio: 1:27 @Sahil, do you think there could be an issue with using gpt-4o when referencing a document which was added before gpt-4o came out? https://cdn.discordapp.com/attachments/1244745971678838895/1245432045161676800/image.png?ex=6658ba7a&is=665768fa&hm=1079e9bb10e4dc0c8ccb1aaf7df3816e9597f0f96111a4286765a0e616974f0e&

Can you please create a new thread along with latest call_id?

Sure

Thank you!

Interestingly, yes, my assistant does use RAG data.