How it works

Every page goes through the same sequence of stages (readanydoc/pipeline/). The model does the reading; ReadAnyDoc does what a model alone does not: it checks the work, finds where every value sits, and returns it in the shape you asked for.

Stage	What happens	Model calls
Intake	Detect the file type, decrypt protected PDFs, convert Office and email files, split into pages	none
Triage	Digital or scanned, rotated or not, is the PDF's own text layer trustworthy	none
Layout	A small open layout model (Docling "heron", ONNX on CPU) finds tables, forms, figures, headers and footers	none
Read	The vision model reads each page at the largest image size it keeps. Scanned pages get zoomed strip re-reads; a character changes only when two zoomed reads agree against the first read. Handwritten pages go to a stronger model.	1 per page, more for scans and handwriting
Verify	Table grids and printed totals are checked without a model; problems become warnings and confidence flags	none
Ground	Every word and table cell is matched to its position on the page (PDF text layer or OCR), block types are set, form fields are paired	none
Assemble	Pages are joined, tables that continue across pages are linked	none
Shape	The output you asked for: blocks, Markdown, chunks, table formats, Reducto's JSON	none

The rules behind it

• The model's text is not changed without independent evidence. A second zoomed read that agrees, the PDF's own text, or a checked total. Everything else is recorded as evidence and confidence, never guessed.
• Never worse than the model alone. Every rule that helped somewhere and hurt elsewhere carries a guard. Ideas that did not beat a plain model read on the full benchmarks were dropped. See How we built it.
• Route by what a page looks like, not by document type. Digital or scanned, tables, handwriting. No templates.

Model-only mode

model-only runs Intake, the single page Read and Shape. No layout model, no re-reads, no second model. See Modes.