ChatPDF

@joel_diigo Thanks! Sure I can share! When uploading the PDF, ChatPDF is using the embedding API from OpenAI to create a semantic index of the different paragraphs of the PDF file. Later when asking a question, the question is also embedded into the same vector space using the embedding endpoint again. This allows us to find the paragraphs that are related to the question from the PDF using a vector similarity metric. A few of these paragraphs are then put into the context of the text-generation API (The ChatGPT API in this case), in addition to a prompt that tells it to synthesize an answer out of these paragraphs. The project you linked is not related, thanks for sharing! There are quite a few projects similar to this. What sets ChatPDF apart is the speed and ease of use :)

@joel_diigo @xathis that is a really interesting approach! Am I right that upon uploading you parse the file and then you get the most similar paragraphs for each question and then feed this context to ChatGPT API for completion?

@joel_diigo @fedoskin yep, exactly!

@xathis How did you create the initial summary and proposed questions? I'm not sure which paragraphs were used as prompts to generate them. Anyway, the implementation is impressive and the output is really excellent. You are the best!

@neofrigate Thanks! For the initial greeting message, it's just using the first few paragraphs of the PDF

@nikolay_karelin Thanks Nikolay! I'll let you know when I decide to write about it :)

@fedoskin Thanks Artem! To "analyze" the PDF, different chunks are embedded using the embedding API, not a typical text completion prompt. I described more over here: https://www.producthunt.com/post...

@jason_derekis Thanks! Sure, I'd be flattered to be featured on easywithai!