First post :) --- Looking for experts with Assistants API

Hi guys,

I hope you’re well. I’m looking for some experts with the Assistants API. Currently, we are looking at if we should use Chat Completion or Assistants and are also struggling with understanding the costs of both.

If you have some experience with this, please do reach out to me and I will message you my Whatsapp.

I’ve noticed assistants api to be slow for my complex multi modal output use-case (Asking assistant api to generate code, run and capture the output to a chart

we are making an AI chatbot of athletes and personal trainers etc. we want it to capture their personality and chat like how an athlete would. Imagine an AI Lebron, we want it to speak like Lebron. What would you recommend and why wouldn’t assistants be good?

Assistants provides an autonomous agent. You set it loose on a user input, and it can make multiple internal AI calls to fulfill from its knowledge, run functions (including ones you coded), run python code, all by the AI’s own choosing and capabilities (or lack thereof), all without your supervision or control of how much is being spent from your API account. Then after waiting and continuing to check back with network requests, you might find there is an answer ready for you.

Not good for any production that faces chatbot customers.

I made a sports celebrity bot. It is more a demonstration of how to use the non-chat “completions-instruct” models to get a satisfying single answer from an expert and persona. API playground preset.

Chat completions is instead what you would want. Instead of one long unstructured “prompt”, you send messages. The first message would be a “system” role message, telling the AI its identity and purpose. Here is an API playground preset for “chat” mode, where I give a new identity to the AI and it follows the knowledge about itself: https://platform.openai.com/playground/p/WJHfYLcE3KVckehxaIitHScK?mode=chat

The current AI models are very resistant to wholly dedicating themselves to being an unbreakable persona. Done wrong, and the AI pretraining will have “Lebron” not always believing what you told it, symptomatic in the types of answers it gives.

To me, character chatbots are passé and cringe-worthy. You can go to forefront.ai and see some, which used to be even more nonsense-related - who wouldn’t want to chat with AI Elon Musk or Mr Beast?

And I made your idea with 5 minutes of my own AI enhancement - on a platform that is pivoting out of this “silly prompt” space.

Additional:

The AI is already trained up to a knowledge cutoff, Sept 2021, or on the latest “economy” model of GPT-4-turbo, to April 2023. That’s not quite good enough for “where are you playing tomorrow, sports ball man?”, but that is not really the entertainment factor of chatting with a simulated AI buddy.

With the latter model that has a very large possible input (50k words or so at increasing expense), you can place a whole bunch of prompt info like the above. Explore what else the AI knows first.

Your forum keyword for injecting even more knowledge from documents on demand is “RAG” (retrieval-augmented generation, using embeddings-based vector semantic search).

You pay for the data. You have to figure out and manage how much word (token) consumption is being done by each user query. Then price competitively.