Parse

Parse turns a document into typed blocks with positions: titles, text, tables, form fields (key/value), figures, headers, footers and page numbers, plus Markdown and chunks ready for an LLM or a search index.

Python

import readanydoc

result = readanydoc.parse(
    "statement.pdf",
    mode="engine",            # or "model-only"
    pages=[1, 2, 5],          # 1-based; default all pages
    citations=True,           # word and cell positions
    options={                 # any Reducto parse option, as in the HTTP API
        "formatting": {"table_output_format": "md"},
        "retrieval": {"chunking": {"chunk_mode": "page"}},
    },
    timeout=600,              # seconds; raises readanydoc.pipeline.PipelineTimeout
)

Result	What it holds
result.blocks	list[Block]: type, content, bbox (page, left, top, width, height as fractions), confidence
result.tables	the Table blocks
result.markdown	the whole document as Markdown
result.pages	number of pages parsed
result.usage	calls, input_tokens, output_tokens, cost_usd, seconds, by_model
result.warnings	one line per problem: a page the provider would not return, a missing layout model, a failed check
result.raw	the full Reducto-shaped response

Bytes work too: readanydoc.parse(data, filename="scan.png").

HTTP

curl -s -X POST http://localhost:8000/parse -H "content-type: application/json" -d '{
  "input": "https://example.com/statement.pdf",
  "formatting": {"table_output_format": "md", "add_page_markers": true},
  "retrieval": {"chunking": {"chunk_mode": "section"}, "filter_blocks": ["Header", "Footer"]},
  "settings": {"page_range": {"start": 1, "end": 3}, "return_citations": true}
}'

Options

Option	Values	Notes
retrieval.chunking.chunk_mode	disabled, page, block, section, page_sections, variable	Same rules as Reducto; chunks repeat the headings they sit under
retrieval.chunking.chunk_size, chunk_overlap	numbers	For variable chunks; overlap in characters
retrieval.filter_blocks	block types	Filtered blocks stay listed but leave content / embed
retrieval.embedding_optimized	bool	Tables get a model-written summary in embed
formatting.table_output_format	html, md, json, csv, dynamic, jsonbbox	dynamic = Markdown for small simple tables, HTML otherwise
formatting.add_page_markers	bool	[[START OF PAGE n]] markers
formatting.merge_tables	bool	Consecutive tables with the same columns merge
formatting.include	hyperlinks, comments, change_tracking	DOCX, PPTX and PDF annotations
settings.page_range	object, list, ints, sheet names	Out-of-range is a 400 error
settings.document_password	string	Wrong or missing gives 442
settings.return_images	figure, table	image_url crops on those blocks
settings.return_citations	bool	Word and cell positions
settings.return_ocr_data	bool	result.ocr.words / lines
settings.extract_document_properties	bool	Title, author and so on from PDF / DOCX
settings.embed_pdf_metadata	bool	pdf_url: a searchable copy of the PDF with our text layer
settings.persist_results, force_url_result	bool	Keep the result past the expiry, or always return a URL
enhance.agentic	[{scope, prompt}] for figure, table, text	A prompt applied to every block of that type, edited in place
enhance.advanced_chart_agent	bool	Chart data extracted from figures, with consistency checks
spreadsheet.*	clustering, split_large_tables, include, exclude, max_cell_count	Excel and CSV handling

The full field-by-field comparison with Reducto is in Compatibility.

Block types

Title, Section Header, Text, List Item, Table, Key Value, Figure, Header, Footer, Page Number, Comment, Signature. Each block has a confidence of high or low (see Citations and confidence).