Post by Hao Hoang

Your RAG accuracy problem is a parsing problem you never measured. Everyone tunes the retriever. Reranks. Swaps embedders. Rewrites prompts. Almost nobody checks the one step where the damage actually happens: turning the PDF into text. Pick the wrong parser and your tables collapse into flat strings, your reading order scrambles, your equations vanish, before a single chunk is embedded. You can't retrieve structure that was destroyed on ingestion. Here's the uncomfortable part: most teams choose a parser by vibes. "Docling worked last time." "MinerU has the stars." AllenAI just made that indefensible. They shipped olmOCR-Bench: 1400+ real PDFs, 7000+ checks across the things that actually break, reading order, tables, equations, handwriting, multi-column layout, tiny footnote text. On their own benchmark, olmOCR scores 82.4, and ranks 4th. Chandra (83.1), Infinity-Parser (82.5), PaddleOCR-VL (80.0) all beat it. They built the scoreboard that ranks themselves fourth. That's the credible part. The lesson isn't "use olmOCR." It's that "which parser?" finally has a measurable answer for YOUR corpus, and you've probably never run it. The next 10% of RAG accuracy isn't in your retriever. It's in the step you stopped looking at. Which parser are you running in production right now, and have you ever benchmarked it against your own documents?