Skip to content

Modes: engine and model only

ReadAnyDoc ships two configurations. Both are exactly the configurations measured in the benchmarks, and a test fails if either one changes without new measurements.

engine (default) model-only
Page reading Claude Haiku 5.5, high effort Claude Haiku 5.5, high effort
Layout model, block types and positions yes no
Zoomed re-reads of scanned pages yes no
Handwritten pages to a stronger model Claude Sonnet 5.5 no
Table checks, printed totals yes no
Word-level citations yes block level only
Typical cost per page $0.003 to $0.005 $0.001 to $0.002
Typical time per page 20 to 50 s 8 to 28 s
Needs readanydoc[engine] yes no

When to use which

  • engine for anything a person will check or a system will trust: forms, invoices, statements, contracts, scans, documents where you need positions and citations.
  • model-only for large volumes of clean digital text where you need the words and tables, not the layout, and cost or speed matters most.

Measured difference

On the complete public benchmarks the engine is clearly ahead of the model alone on page layout (+23 to +35 points), form question and answer pairs (+5.8), headers and footers (+10.8) and the olmOCR-bench overall score (+2.0); level on tables, plain text and long extraction. Details in Benchmarks.

Selecting a mode

readanydoc.parse("doc.pdf", mode="model-only")

readanydoc parse doc.pdf --mode model-only

READANYDOC_PROFILE=model-only readanydoc serve

Reducto's settings.model also maps: r-1 runs the configured profile, r-1-mini and legacy run its cheaper tier (engine becomes model-only at medium effort). See Compatibility.