NOBODY IS TALKING ABOUT THE MOST EXPENSIVE STEP IN EVERY AI PIPELINE.. It isn't the model. It isn't the embeddings. It's PDF extraction and most teams are still feeding their LLMs flattened garbage.…

What stood out was the reduction in cleanup work. Instead of spending time fixing formatting issues after extraction, the Markdown output was already organized enough to drop into my workflow with…

Quick reminder if you're building RAG systems: The model isn't your bottleneck. Clean input is. PDF → Clean Markdown → Chunking → Embedding → Retrieval Each step compounds. Bad parsing at step 1…

SPENT TOO LONG CLEANING PDF OUTPUTS BY HAND Found @nutrientdocs ’s CLI, one command and the Markdown comes out clean enough to feed straight into a pipeline. Useful if you work with PDFs and LLMs.…

Where it really shines: text-heavy technical documents. Research papers, documentation, dense reports, it preserves tables, headers, and citations better than anything else I've used. For that use…

Everyone is obsessed with better RAG systems. Almost nobody is fixing the first step. Getting clean data. Because AI doesn't read PDFs. It reads text extracted from PDFs. And that's where most…

Would really re-run this on a commonly used open-source benchmark like opendataloader's that has all these parsers on there - I ran it and the speed increase is impressive (nice job on the Rust…

The newly launched Data Extraction API is powered by our enhanced ICR pipeline, which leverages some clever segmentation and specialized modes for parsing document elements. We ran it through the…