ReadAnyDoc

Parse and extract documents with vision models, on your own machine and your own model account. Open source, Reducto-compatible, built by Unmute. Website: https://arpitpatel25.github.io/ReadAnyDoc/

ReadAnyDoc reads every page of a PDF, scan, photo or Office file with a vision model (Claude Haiku 5.5 by default) and returns typed blocks, tables, Markdown, or the fields of your schema, each with the place on the page it came from. Around the model sits an engine that adds what a model alone does not give you: block types and positions, form fields, headers and footers, table checks, word-level citations. It runs as a Python library or as a server that speaks Reducto's HTTP API, so existing Reducto integrations work by changing one URL.

import readanydoc

result = readanydoc.parse("invoice.pdf")
print(result.markdown)

Results

Five public benchmarks (about 4,250 pages and 10 long documents), each scored with the benchmark's own grader (no AI judge). Same documents and same requests for all three systems. Details and caveats: BENCHMARKS.md.

Where Reducto is ahead: page layout (by about 6 points), formulas, header and footer removal, and speed (about 5 s a page against 20-50 s here).

Install

pip install "readanydoc[engine]"        # the engine: layout model and OCR run on your CPU
pip install readanydoc                  # model-only mode, no local models

The engine also uses Tesseract when it is installed (brew install tesseract / apt install tesseract-ocr). The layout and OCR weights (about 200 MB, Apache-2.0) are downloaded on first use, so the very first parse takes a few minutes longer; later runs start straight away. Python 3.12.

Set your model provider's key: export ANTHROPIC_API_KEY=... (or OPENAI_API_KEY for OpenAI models).

Use

Parse

result = readanydoc.parse("statement.pdf")
result.blocks          # [Block(type="Table", content="<table>...", bbox=BoundingBox(...)), ...]
result.tables          # the Table blocks
result.markdown        # the whole document as Markdown
result.usage           # model calls, tokens, cost in USD, seconds
result.warnings        # anything that went wrong on a page

Extract into your own types

from pydantic import BaseModel

class Invoice(BaseModel):
    number: str
    issued: str
    total: float

out = readanydoc.extract("invoice.pdf", schema=Invoice)
out.data                      # Invoice(number="INV-2041", issued="2026-03-01", total=1575.0)
out.citations["total"]        # [Citation(content="1,575.00", bbox=BoundingBox(page=1, left=0.81, ...))]

A JSON Schema dict works as well as a pydantic class. instructions= is the system prompt for the extraction model; array_extract=True processes long documents in parts and merges their lists.

Choose the mode and the models

Every model the pipeline uses can be chosen, with its effort level and any further request options:

readanydoc.parse(
    "form.pdf",
    mode="engine",
    model="claude-haiku-5-5", effort="high",                 # page reading (the default)
    model_options={"max_tokens": 32000},                     # sent as given to the provider
    handwriting_model="claude-sonnet-5-5",                   # handwritten pages (the default); None = like any page
)

readanydoc.extract("contract.pdf", schema=Terms,
                 model="claude-haiku-5-5",                   # reads the pages
                 extract_model="claude-sonnet-5-5", extract_effort="high")   # fills the schema

Models: any Claude model, OpenAI models (gpt-...), or any OpenAI-compatible endpoint written as openai_compatible:<model>@<base_url>. The benchmark numbers above are for Claude Haiku 5.5; other models work but have not been measured here.

Command line

readanydoc parse invoice.pdf --markdown
readanydoc parse scan.pdf --mode model-only --model claude-sonnet-5-5
readanydoc extract invoice.pdf --schema invoice.schema.json

Server (Reducto-compatible API)

readanydoc serve --port 8000
# or: docker compose up

from reducto import Reducto   # Reducto's own SDK, unchanged

client = Reducto(api_key="anything", base_url="http://127.0.0.1:8000")
upload = client.upload(file=Path("invoice.pdf"))
result = client.parse.run(input=upload.file_id)

Parse, extract, split, classify, edit, async jobs, webhooks and citations are implemented. Configuration, security settings and a route-by-route compatibility table: docs/server.md.

How it works

Each page goes through a fixed sequence of stages (readanydoc/pipeline/): intake (file type, pages), triage (digital or scanned, rotation), layout (an open layout model on CPU), read (the vision model; zoomed re-reads of scanned pages where two reads must agree before a character changes; handwritten pages to a stronger model), verify (table grids and printed totals checked without a model), ground (word and cell positions, block types, form fields), assemble and shape (the requested output format). Nothing the model wrote is changed without independent evidence. Design decisions and their reasons: docs/decisions.md.

Limits

• Slower than Reducto: 20-50 s per page here against about 5 s.
• About 1.5-3% of pages in large public benchmarks come back empty because the model provider declines to return their text (content filtering); the response says so in warnings.
• Confidence values are tiers from how a value was found on the page, not calibrated probabilities.

Documentation

Contributing

See CONTRIBUTING.md. Security reports: SECURITY.md.

License

Apache-2.0. The layout model (Docling "heron") and the PP-OCR weights are Apache-2.0.