Differences in API and ChatGPT end user app

I am working on a project for my company. My testers are complaining that what I provide them is not a good alternative to end use app (ChatGPT)

I am using model chatgpt4o

Missing/different items:

  1. API does not answer correctly to the questions like “what date is it”. It does not know the actual date. How can I adjust that other than giving an “instruction” parameter like “today is …”. What is the official method ?

  2. API does not answer correctly to the questions like “how is the weather”. Of course it doesn’t know the location of the user. Would it provide correct answer if I provide an instruction parameter with user’s location ? Is there a suggested method for this ?

  3. I am using speech to text endpoint for microphone input. There are some words or sentences similar in sound in different languages. For example “merhaba” is same in Turkish and Arabic. When the user is talking Turkish most of the time endpoint understands in Turkish, but in between sometimes endpoint understands it as if Arabic. I don’t want to enforce the language input from user’s locale. Because then when the user wants to use another language, it will not be interpreted correctly. Will it ? What is the best approach here ?

  4. On API , we have 6 voices to use for text to speech. But on the app there are 5 voices and some of them are really better than API voices. Will they be added to the API later ?

If using “Assistants”, you don’t have 100% control over the input that is sent every time.

There is an “additional_instructions” parameter that you can include, that would be the logical place remaining to place a computer-generated date, and then improve that with the locale so it switches for the user.

AI is not a personal assistant that will answer question such as “play the latest Taylor Swift song”. It can call on programmed tools you provide by function-calling if you need a weather bot, though, using some separate service that provides that information. All its knowledge is pretrained in a long process, and fixed unless you input more text when you call the AI.

As programmer, you can give a user interface with multiple remembered settings, such as the language the user wants to speak and have recognized. Then it can be passed by ISO code to whisper.

The voice actors seem to have signed particular agreements for how their voice would be used, and OpenAI wants to keep their ChatGPT separate from the products of another. So the voices seem like they would remain separate, as their is no technical reason for the current separation.

1 Like

@_J thanks for your comment.
Actuallly, the problem is not what I can do or can not do on the user interface.
I opened this thread to discuss the differences between API and the ChatGPT app.
For example, I don’t understand the reason why API does not know the current date while ChatGPT knows.
Of course we can use instruction parameter , it is very simple. But I asked it here because I might be missing something. Maybe a misconfiguration I do while creating the assistant and/or thread.
And API really has too few female voice options compared to the app. That’s not fair…

Use tools for these (“function calls”).

https://platform.openai.com/docs/assistants/tools/function-calling

1 Like

ChatGPT knows because OpenAI provided it a system message to let it know.

You are ChatGPT, a large language model trained by OpenAI, based on the GPT-4 architecture. Knowledge cutoff: 2023-12
Current date: 2024-05-16

ChatGPT now has one less female voice…

Here’s the Nova voice, where I previously showed what having the AI write some thinking pauses would sound like.

I could not understand what you meant here. Is the Plus version acting similar to what API provides ?
ps. English is not my native language. So maybe you wrote something meaningful but I did not understand :slight_smile:

I can assure you: nothing meaningful was written.

1 Like

how come ?
on API which voice do you accept as female ?
I only see nova and shimmer. So 2 of them
On the app, we have Juniper, Sky and Breeze (3)

Sky has been removed due to controversy. It’s now just one of the other voices.

I hadn’t noticed it. interesting…

Prior to calling the API create a var with the date and location of the client. Then concatenate the var along with your call to the API.

The voice issue is separate deal. I haven’t played with this feature.