I finally built a Supercharged PDFChat and now constantly improvising based on feedback of 1000+ users!! GPT-4 32k supported.
Being an NVIDIA AI Inception member, I have been working on solving most of the pain points of these Chat PDF Tools. The problems that hurt the most was inaccurate retrieval, performing any sort of task along with retrieval, and handling large PDF’s.
Most chat pdf tools cannot:
handle large PDF’s, unstructured PDF’s
break while performing tasks like “explain me like I am 5 wrt to … in the pdf” along with retrieved content .This is because the prompts used are pretty standard and have not been improvised.
does not work well with scanned PDF’s.
cannot summarise any document longer than context window.
Well what did I do differently?
made a custom OCR to extract most of the text from any PDF thereby increasing the quality of retrieval.
made two modes two modes for querying. One is Strict QnA, one is ASK AI, the former one for only question answering tasks grounded in the information in the document(least hallucination), and the latter for performing tasks with retrieval with support from Internal Knowledge base of GPT.
It also works with most languages, same as gpt api(90+ languages)
When it comes to summarization, most chat pdf tools will only summarise a part of the document, but I use a specific approach to this problem which involves clustering. BY this method I am able to summarise PDF’s longer than 1000 pages.
Oh, PDFCHAT also has a special feature to convert any scanned invoice to CSV or JSON on the go.
–And a multi document feature which allows to compare information across two PDF’s simultaneously (still improvising on this but for semantically different content it works pretty well).
We offer a generous free trial and it costs 12 bucks a month to have access to a GPT-4 PDF Chat app. Please try it out and leave some feedback! Thanks!