I’d like to use the new 3.5 turbo model if it’s better, faster and/or cheaper. However, my application is NOT a chat application. It’s an embeddings-driven search and Q&A application. I don’t want a rolling Q&A chat because my text is dense and complex and I need to maximize the number of tokens from the search results that are fed into the prompt, for maximum context. It doesn’t seem right that non-chat applications should have to be re-worked in the back end as chat-style applications. It’s counter-intuitive to create the “roles,” not to mention the wasted time wrangling my code into a new format. It took me a very long time and lots of experimentation with prompt engineering to get where I am. How hard can it be for OpenAI to offer the upgraded model in a non-chat version? ChatGPT has created huge hype, but it seems like OpenAI has forgotten that many customers are not riding the chat bandwagon. More generally, isn’t it good practice to release model updates that integrate seamlessly with the work that developers have already done? I think we should be able to simply substitute the model name in our code. I’d appreciate knowing if OpenAI plans to offer this soon, or whether I really have to re-work my code as if I’m building a chat application.
turbo model is not for embeddings so there is zero reason to consider or migrate to
End of story. Stick with your current embeddings model.
I use text-davinci-003 in the second step – question answering – that’s what I’m frustrated about.
The model above is also not an “embeddings model”, it’s a “completion” model, isn’t that right @lmccallum ?. The main embeddings model I always use is:
Sorry, @lmccallum I have read your post twice and still cannot figure out what you are posting about.