@barosell/streamllm/src/ai.ts

Support streaming LLM responses using rxjs observables

import OpenAI, { ClientOptions } from "openai";
import { ChatCompletionMessageParam } from "openai/resources";
import { Observable, Subject } from "rxjs";

export type ChatRole = 'SYSTEM' | 'USER' | 'ASSISTANT';

export interface Completion {
  role: ChatRole;
  content: string;
}

export class AiInterface {
  constructor(
    private openai: OpenAI,
    private model: string
  ) {  }

  stream(completions: Completion[]): Observable<string> {
    const deltas = new Subject<string>();

    (async () => {
      try {
        const stream = await this.openai.chat.completions.create({
          messages: completions.map(c => ({ role: c.role.toLowerCase(), content: c.content } as ChatCompletionMessageParam)), 
          model: this.model,
          stream: true,
        });

        let content = '';
        for await (const part of stream) {
          content += part.choices[0]?.delta?.content || '';
          deltas.next(part?.choices[0]?.delta?.content || '');
        }
  
        deltas.complete();
      } catch(error) {
        deltas.error(error)
      }
    })();

    return deltas;
  }
}