MinerU

MinerU converts documents that were written for humans into something a model can read: PDFs, Office files and images in; structured Markdown with layout, reading order, formulas and tables preserved. It runs as a library, a CLI or a service, and it can be pointed at different backends — small models on ONNX or Torch, or a VLM served through llama.cpp, vLLM or LMDeploy. It also keeps page and block locators so a downstream agent can cite where a statement came from.

This is the unglamorous stage where retrieval quality is decided, and PDF is why. A PDF is a description of where ink goes, not what the text means: two-column papers, tables whose rows span pages, formulas as glyph soup, scanned pages with no text layer at all. Every chunking strategy downstream is built on whatever this stage decided, and a table split mid-row is not recoverable by any amount of tuning later — which is why parsing deserves the attention usually spent on the embedding model. The caveat is that quality is domain-specific: a pipeline tuned on academic papers meets an invoice or a scanned form and behaves differently, so the test is your own document types.