Every RAG eval deck I’ve seen obsesses over the reranker and the embedding model. Almost none of them measure the step that bounds all of it: how faithfully the document got parsed in the first place. Baidu’s Unlimited-OCR — explicitly built to push DeepSeek-OCR further — aims squarely at that front of the pipeline, with a pitch of “one-shot long-horizon parsing.”

• 📄 Whole documents, one pass — multi-page PDFs parsed in a single 32K-context shot instead of page-by-page stitching, which is exactly where layout and table structure usually get mangled.
• 🔁 Repetition collapse handled at decode — a no-repeat-ngram logit processor (size 35, large window) fights the degenerate looping that plagues long-document OCR transformers.
• 🎯 Two render modes — a cropped “gundam” config for dense single images, a base config for multi-page runs; you’re tuning input resolution, not just prompting.
• ⚡ Built to serve, not just to demo — a transformers path plus an SGLang server with an OpenAI-compatible streaming API and a batch infer.py. That’s an ingestion worker, not a notebook cell.
• 🔍 The eval gap stays open — the HN discussion wants apples-to-apples numbers against DeepSeek-OCR and the usual baselines; the README leads with code, not a benchmark table.

If your retrieval quality is capped anywhere, my money’s on it being right here — garbled tables and dropped headers that no reranker downstream can recover. The real question for anyone running document RAG in production: when did you last put a number on parse fidelity, separate from retrieval recall?