A typical earning report will be a pdf file consisting text, images(charts usually) and tabular content (I actually think most tables you see in the pdf are images anyway).
My goal is to extract all the text and images and feed to ChatGPT to analyse.
What’s your best practice chunk strategy here?
I am going by one page one chunk, assuming most single page in any financial report is unlikely to exceed the token limit of 4096.
I also can’t upload the images to GPT-4 yet, so I could only extract the text out of images and combine with the text extraction for each page. Then feed each chunk (one page content) to ChatGPT to summarise and rejoin all the summaries in the end. Maybe summarise it again.
Just want to hear how you amazing people would tackle this.