Parse PDFs at lightspeed (this video is at 1x) Absolute cinema

Parsing PDFs is not that difficult if you don’t obsess about the format. Problems appear when the PDF is not OCR’d and is some random image taken from some random app but in PDF format. I’m not sure if liteparse fixes that

hey i think there's a bug in your bbox generation for pages with rotated text columns (like the arxiv watermark). i just tried it with random papers and it extracted text for the first page but the bboxes are all offset.

ooh cc @LoganMarkewich for visibility

Unfortunately rather useless if it cannot output into markdown preserving structure.

we are working on markdown

Would also be interesting if you could benchmark against @firecrawl pdf-inspector, also built in rust

Nice work @jerryjliu0 and team. So snappy.

@llama_index is it open source?

@llama_index yes!

This is amazing

Brilliant work! We’ll definitely give it a go and possibly replace our existing custom parsing logic with it!

Honestly, this is the sort of speed that changes the feel of agents. A model-free parser ripping through a resume at 1x is funny because the model finally gets to do the clever bit instead of waiting for page 47 to become text.

But the tables bro, what about the tables?

Pain resolved😭

Interesting

Did it support arabic language?

Does it support RTL languages? I tested it on a Persian/Farsi PDF file and the output text wasn't quite right.

Can we just stop producing PDFs :)

@grok with my Opus4.8 kiro-cli agent is it better to just let the model parse a pdf or let the model use lightparse v2 as a tool. Frontier models can parse PDFs themselves but these tools might offer better parsing than the model. Do research, think hard, take your time.

I triex it.. It working good but is there any way to convert it into html like docling or open data loader?

the speed difference in the video is wild. I've been waiting for a parser that doesn't make me wait

the model-free angle is the whole game. parse extraction that doesnt depend on a vendor is exactly what agent workflows need to be reliable.

Impressive, is the output markdown?

looks amazing, thanks a lot!

brb watching this on loop, that pdf parse is clean

Going to save so much time with this

How does it do for pdfs created through tools like AutoCAD? The fact that they contains diagrams mostly