Customize the Adapter to be able to send conversation history in api call for openai

[whisper-ai-team]

Hi @salmenus, I have a requirement to send the conversation history in open-ai chat completion request. Is it possible to achieve this with custom adapter settings?

[salmenus]

hi @whisper-ai-team - Yes, it is possible.

If you look at the extras parameter passed to custom adapters. It has the conversation history.

Doc here: https://nlux.dev/api/adapters/custom-adapters

[whisper-ai-team]

Thanks great!

1. Is it possible to change this history types, as open-ai expects in this format
   {
   content: ""
   role: "user"
   }
   openai expects these roles: 'user' | 'system' | 'assistant'

current defined format in library

type ParticipantRole = 'user' | 'system' | 'ai';

type ConversationItem = {
role: ParticipantRole;
message: string;
};

2. And also is their any provision to manually stop generating the response same as we have in chatgpt?

[salmenus]

• Regarding the roles: You can do the mapping yourself to whatever format works best.

extras. conversationHistory?.map((item) => {
    return {
        /* Transfer each item to whatever format works for you */
    };
});

nlux is made to work with multiple LLMs and backends, and not only ChatGPT, and thus we cannot only optimise for ChatGPT API

• Q2: What do you mean by "manually stop generating the response" ?

If you're looking for fetch the response in a single HTTP request, you can use fetch adapter instead of streaming adapter. You can also change the animation speed when the response is being generated param here.

[whisper-ai-team]

Thanks, for now Iam using mapping.
"manually stop generating the response" - > I mean in chatgpt there is a noption to stop the on going response stream, please find this sdk hook reference - use-chat - vercel AI SDK, which offers stop method.

[salmenus]

Ok. Got it. Thanks for explaining.

Yeah, we don't have that option yet, but we can probably add it.

I added a ticket in the project backlog here to track this feature request.