PDF documents with both images and text for a RAG ...
# 🧶︱generalc
cuddly-rain-31144
09/10/2025, 3:57 PMPDF documents with both images and text for a RAG system. I’ve been using docling for parsing (https://github.com/docling-project/docling) and works quite well.
m
mammoth-lighter-94680
09/10/2025, 4:03 PMFun fact Pinecone Assistant supports multimodal documents now!https://www.linkedin.com/posts/pinecone-io_public-preview-multimodal-context-for-activity-7371180691696308224-aDod?utm_source=share&utm_medium=member_desktop&rcm=ACoAAAnDaFsBjjsIMX5jT1mncWaYE2TKs2ecN-4
mammoth-lighter-94680
09/10/2025, 4:03 PMYou can do your parsing and search and generation just with Pinecone if you wanted to!
That might help with latency issues related to docling. I remember trying it and feeling it to be slow
mammoth-lighter-94680
09/10/2025, 4:03 PMOtherwise, I've heard good things on Cohere's model as well!
h
handsome-oxygen-95530
09/10/2025, 4:04 PMAmazing !
m
mammoth-lighter-94680
09/10/2025, 4:20 PMc
cuddly-rain-31144
09/10/2025, 6:24 PMWait so does this mean that the parsing doesn’t have to be done by us anymore? Pinecone can handle it?
cuddly-rain-31144
09/10/2025, 6:30 PMcuddly-rain-31144
09/10/2025, 6:32 PMJust for clarification, for the time being the PDF limit is 20, In my case I have a couple hundred PDFs so is there any expected updates to that figure?
cuddly-rain-31144
09/10/2025, 6:32 PMSorry for the bombardment of questions 😂
m
mammoth-lighter-94680
09/10/2025, 8:18 PMYes, assistant chunks and parses PDFs for you
mammoth-lighter-94680
09/10/2025, 8:21 PMNo problem. We'll probably expand the number in the future but I'd have to consult our product team on the exact vision. In the meantime, I suggest testing on a small scale first and let us know when you start hitting this limit.
One workaround might be to combine PDFs temporarily until you hit the 100 page mark
c
cuddly-rain-31144
09/10/2025, 10:35 PMOk sounds good, was just thinking of that solution! Will definitely try that out. Thanks a lot Arjun. 🙏🏼
cuddly-rain-31144
09/10/2025, 10:35 PMIs there any indication as to what sort of chunking is performed? Like is it semantic? Fixed? Or something else
m
mammoth-lighter-94680
09/11/2025, 3:53 PMWe don't have anything public facing on exactly how we do this just yet, but stay tuned 🙂 The feature is early and in public preview for now.
c
crooked-fireman-72082
09/11/2025, 4:48 PM@cuddly-rain-31144 Curious to hear about your docling experience, we found it good for prototyping not production ready yet for scale based on evals. @mammoth-lighter-94680 - sounds like you guys are on it to improve the doc processing pipeline.
c
cuddly-rain-31144
09/11/2025, 5:09 PMAlright no worries. Thanks.
cuddly-rain-31144
09/11/2025, 5:35 PMHey Kavita, like you said, in my opinion its not ready for production just yet but unfortunately it might be the best option available at the moment. Luckily i don't need it to be particularly good at scaling since I'm using for a small application. Let me know if you find a better alternative!
c
crooked-fireman-72082
09/11/2025, 5:48 PMLandingAI's doc extraction or LangExtract by Google - good and better performing alternates to docling and diy with a really good prompt definition and tools. Depends on your contents in the doc - tables, charts, visuals..
c
cuddly-rain-31144
09/12/2025, 5:11 PMAh nice, thanks a lot.
cuddly-rain-31144
09/12/2025, 5:11 PMWill have a look at those then
3 Views