I’m having trouble getting the outputs I need from my current workflow. I’m taking long interviews (about 10,000 tokens) chunking and summarizing the with text-davinci-003, then recombining it into a single 900 token (approximate) summary. I then ask the model to write a story about the subjects life about 4-5 paragraphs long “using the following information”. If I pass this prompt with summary to chatGPT I more or less get what I want, but often times I’ll get nothing from the api or just a single sentence. Is there any trick to getting text-davinci-003 to create a story from a large amount of data that you present in the prompt?
Unless you are trying to summarize content that may contain items that would flag content restrictions, I think you should instead use gpt-3.5-turbo-16k which should have more than enough tokens to handle both the ingestion of the data and the response.
If you really need to use davinci then you will want to take chunks of text, summarize them into some succinct points and then create a summary of the chunks of text. But this is very hit or miss compared to using the model that has enough tokens to handle what you are attempting. Sometimes though especially with content that is a bit more fringe for Open AI Davinci might be the best option you have.