I'm looking for a faster way to load PDFs than pymupdf4llms.

Question

I'm trying to find the "best" way to load PDF's, and I settled for now on the pymupdf4llms because I was using the PdfMarkdownReader from the marker-py lib, but I was hoping to avoid insanely long indexing time

I think SentenceSplitters requires torch, which makes sense, but on any decent server they probably take a long while I presume if it goes at 2.19it/s for my computer with a 4090 24 GB?

Logan M · Answer

sentence splitters do not require torch?

ZachHandley · Answer

oh, then I guess pymupdf4llms or something? or the pptx reader actually

ZachHandley · Answer

@Logan M is it possible to use PptxReader without torch

ZachHandley · Answer

it's killing our other devs ability to use it as it's so long to install it literally is stopping them lol

Logan M · Answer

the pptx has hard dependencies on transformers (its using them to extract content)I recommend writing your own reader?

ZachHandley · Answer

so

ZachHandley · Answer

I basically copied yours but used pptx to extract text and stuff, and then an LLM to see the page

ZachHandley · Answer

I think it would be beneficial to offer that to users too, cause the only change required would be using a vision LLM to "see" what's on the page, vs using torch

ZachHandley · Answer

updating the pptx one still to remove some deps

ZachHandley · Answer

cause it was a custom one

Find answers from the community

I'm looking for a faster way to load PDFs than pymupdf4llms.