Choosing an architecture for a client-only facing cuatom GPT

API
1

HI,

I’m not sure this is the right place to ask, but…

I’m trying to figure out the best approach to doing something for a client. He wants to create a Web page where users can use his Custom GPTs. He only wants his clients to have access, so he wants to make sure that they have an account with him before letting them use the GPTs.

Second, his GPTs need RAG to be usable. He will be providing them with many documents to parse and get appropriate answers.

Ideally, he would like not to force his clients to buy a Plus prescription to use the GPTs as needed (i.e., he doesn’t want them to run into the free GPT limit).

So, the options I see are:

  1. Tell the clients they need a Plus or Teams account to use his GPTs. On the plus side, it gives clients unlimited use of his GPTs. On the down side, it’s harder for him to provide a universal interface to access those GPTs. That said, I’ve seen this done and I think it’s acceptable, but he may not. There’s also the authentication aspect (I’ve seen the discussions on using OAuth servers for this, so that might be a solution).

  2. Create an assistant and call it from the website through a webhook. It could be connected to a Make/Zapier/other type of solution that does authentication . The challenge becomes managing API call costs because the RAG information will be required on every call. I tested, and two calls were 35 cents (with 4o), which is a pretty steep price, I think. Are there efficient methods I can use to reduce that cost? (I haven’t started looking into it yet).

  3. Fine-tuning with 3.5. There’s an upfront cost, but after that’s done, the cost per call is significantly less than 4o. I would have to determine at what point it becomes less costly than just a straight call to 4o.

Any thoughts? Is there something I’m not groking or that I’m overlooking?

Thanks,

L

2

Number 1 is not a workable solution at all. There is no concept of exclusive sharing of a GPT outside of Team or Enterprise accounts, so unless they plan on providing a Team or Enterprise account to each client it’s a non-starter.

Number 3 isn’t a different option as much as a variation of Number 2.

The only way to accomplish what they want to do is by building a custom application with the API. That can be either with an assistants or with a chat/completions API, but it needs to be through the API and they’ll need to pay for each exchange.

3

Thanks, that’s pretty much what I figured.

L