11–15 of 15 posts
What is this? The ability for chatGPT to parse text in PDFs? Couldn’t it already do OCR on text in images?
Very recently it got vision which can be used for OCR on a single image, but that's massively inefficient and limited compared to what this is likely doing for longer documents
Earlier quoted context omitted.
Very recently it got vision which can be used for OCR on a single image, but that's massively inefficient and limited compared to what this is likely doing for longer documents
I tried using the vision feature for OCR & it was worse than Tesseract. At least for financial documents where you need exact numbers amounts. Will the new PDF feature be better? I’m not so hopeful.
There is practically no chance the new feature uses vision because that'd be _insanely_ slow and expensive for any reasonably sized document. They're likely using Azure's LayoutLM derived tech to get out text, then using embeddings to answer on questions
Admittedly, I don't know how the implementation works, but I was expecting it to be able to do a search on the pdf to find the relevant parts and answer requests, but my results have been really bad. It makes up stuff seemingly more when I ask using a pdf than when I ask it without a reference.
I guess they are rolling it out gradually because mine hasn't changed. At least not on mobile web. Haven't looked at desktop.
This pdf feature seems very useful, but there's no instruction on what it's doing under the hood or how to use it best.
Earlier quoted context omitted.
I tried using the vision feature for OCR & it was worse than Tesseract. At least for financial documents where you need exact numbers amounts. Will the new PDF feature be better? I’m not so hopeful.
Using the vision feature for OCR is like using an LLM for math: it might work, but we already have a lot of tools that are hyper-optimized for the task. There is practically no chance the new feature uses vision because that'd be _insanely_ slow and expensive for any reasonably sized document. They're likely using Azure's LayoutLM derived tech to get out text, then using embeddings to answer on questions