Show HN: Chat with your PDF (my first AI side project)
It's a simple web app to help you clarify long complicated PDFs by asking it questions. How it works:
- Upload a PDF - Ask it questions - It gives you answers, based on the PDF, generated by OpenAI's GPT-3.5 model.
Yes, it's not new, but with mine, there's no subscription of signup. Just a one time $1.99 payment per doc. So yes it is a paid app (OpenAI is not free) but here's a 100% discount code so you can try it out:
PDFS4HN
If you do, I’d love feedback. The product is embarrassingly basic and the prompts need a lot of work, so if you could comment with your feedback below, or email to alfarez@gmail.com, I’d highly appreciate it!
Background:
This is my first AI app which I built as a way to learn AI. Well, I did start learning without a project in mind, but boy the number of rabbit holes you could go deep into is too much. So I thought building something would help me guide my learning. After all, the goal is to build something with the knowledge.
A pleasant side effect with trying to build and launch a product is that I did not just learn the tech, but also all the other important stuff about engineering an AI SaaS, what the UX for an AI backend might look like, and how to talk to the target users of the product without putting them off with all the AI jargon.
Because we all know, at the end of the day, nobody cares about your tech. They just care about how it can help them make money, save money or fix a headache!
Look forward to have you all try it out!
5 comments
[ 3.2 ms ] story [ 18.2 ms ] threadBut you can upload up to 10MB in file size.
That UX flow seems strange because the screen flow should just be: 1) upload, 2) pay, 3) chat. It shouldn't have asked you to check your email for anything.
Please feel free to email me at alfarez at gmail dot com if you're still seeing this issue.
As for handling large files, I'm using a popular method of breaking down the doc into small chunks and then just sending the relevant chunks to the AI.
How to pick the "relevant" chunks? Use embeddings and search for the embeddings that are "semantically closest" to the question being asked.
I wrote up an outline of this process here: https://farez.me/how-to-create-a-chatbot-with-openais-api-a-...