Four projects tracked here all turn documents into Markdown and all get recommended for RAG. They solve different problems and the licences differ, which matters more than the benchmarks.
Docling (IBM, MIT). Layout analysis plus table structure recognition; exports Markdown, HTML and a lossless JSON document model; has integrations for LangChain and LlamaIndex. The strongest when tables matter and when you want a structured representation, not only text.
MinerU (OpenDataLab, AGPL-3.0). Aimed at scientific PDFs: formulas to LaTeX, multi-column layouts, figure and caption handling. Heavier model stack. AGPL is fine for an internal pipeline and a conversation for anything you distribute.
Marker (Datalab, GPL-3.0 for the code, with the model weights under additional terms that restrict commercial use above a revenue threshold). Fast, good general-purpose Markdown output, OCR fallback. Read the weights terms before building a product on it.
markitdown (Microsoft, MIT). The simplest of the four: converts Office files, HTML, PDF and more to Markdown for LLM consumption. Not layout-aware. For Word and PowerPoint input it is the pragmatic choice; for a two-column scanned PDF it is the wrong one.
Unstructured (Apache-2.0) is the fifth name people compare, more of an ingestion framework than a converter.
Pairs we have written up: /compare/docling-vs-mineru, /compare/marker-vs-mineru, /compare/docling-vs-unstructured. The short version: Docling by default, MinerU for papers, markitdown for Office files, Marker when speed on general PDFs matters and the licence fits.