https://www.boundaryml.com logo
Powered by
hey all has anybody successfully used
# general
r
hey all has anybody successfully used BAML for a document parsing pipeline ?
a
wdym by document parsing? Extract tabular content?
b
This is a super common use case! What are you trying to do specifically?
r
Docling, marker, GoogleDocument, Tesseract etc. provide complete parsing solution . I was wondering if anybody here tested a custom workflow using vision models/convertion pipelines with BAML for the restructuring of the data
e.g. : PDF -> a) vision models extract data or b) convert PDF to markdown with e.g. PymPDF4LLM -> result is usually 80% usable -> BAML used for extracting structured outputs
maybe a single multi-modal model with BAML could already provide a complete structured output of a PDF page (images and tables would be extracted seperately)
a
Version a is what I use. I recommend using Gemini 3 pro as it consistently performs better than any other model with vision
b
We’ve got customers who use multimodal capabilities with all the major model providers And as you point out, for some tasks you can one-shot them using BAML
eg one of the challenges with an extraction approach is that a markdown extractor assumes your image/PDF is easily represented with markdown: what if the document uses different font sizes in addition to underline/italics/bold, or uses column colors to assign meaning to different fields?
r
yes good point, downside is that even with BAML you would need a very good model that can recognize and write LaTeX formulas / tables etc. @anvith thanks will give it a go, hoping gemini-3-flash is able to deliver.
PreviousNext
Powered by