----------------------------- **While I generate an answer, here are some sources that might be helpful**: * [Set up your Server URL]() * [Call Functions]() * [Billing Examples]() * [Inbound Support Example]() -----------------------------

### Answer I understand your concerns about the performance of models, including instances of hallucinations. Vapi integrates with OpenAI's models, offering flexibility in model usage and function calling to enhance performance and reliability in various scenarios [(1)]() . For specific issues or optimization advice, it might be beneficial to explore Vapi's documentation or contact support directly for tailored assistance.

using the same exact prompt on both

AZURE open ai is just dumber

" AI: Great. Let's look at Tuesday. For Tuesday, April 20 third, we have a wide range of times available. We're open from 9 8 to 4 acid. Have a preferred time in the morning or afternoon? User: 9 what to for what? AI: My apologies for the confusion. On Tuesday, April 20 third, we have openings from 9 at to 4 Aje. You prefer a morning or an afternoon appointment? User: Uh, afternoon? " there is hallucination and then there is this.

"from 9 8 to 4 acid" it seems like its on acid

@User GPT 4 would be too expensive, could you suggest any model that doesn't trip like this and is reliable?

have you given any thoughts on this?

I was thinking the same

what about sonnet?

and why haiku, isn't it the smallest model?

and do function calls work as well with all of these?

I will test all of them but would appreciate some pointers from people who have already tested

from what you have shown, i believe you should make the instructions shorter. The descriptions for the functions on the functions panel already do a lot by itself

if you are talking about system prompt, when you give it general instructions it was always messing it up cuz its too dumb. so I had to give it specific example for each scenario, my prompt is like 1400 tokens

no matter how many times I say don't just dump all the slots as a list, it would still do it

its prone to print out information then explain and summarize it

so examples are needed to change its behaviour

8888888888888888888888888888888888888888

User: Hey, Melissa. How's it going? AI: Speaking in a friendly professional tone, hello? This is Melissa from California Dental. I'm doing well. Thank you for asking. How can I assist you today? Do you have a dental concern or would you like to schedule an appointment. User: Right. What? No. The first thing is tell me how are you doing. AI: Chuckles warmly, oh, I'm doing splendidly. Thank you for asking. As an AI assistant, I don't have personal feelings, but I'm here and ready to help you with anything related to scheduling a dental appointment. How can I assist you today? User: Yeah. Well, I have a tooth issue. Like, uh, pain? haiku is all sorts of wrong

also didn't call the function and hallucinated that it did

I know there is a bunch which one would be best for less hallucinatioon and good function calling?