Extract¶

Extract fills in a schema you define: a pydantic class or a JSON Schema. You get your own type back, with a citation for every value.

Python¶

from typing import Optional
from pydantic import BaseModel
import readanydoc

class Party(BaseModel):
    name: str
    role: str

class Contract(BaseModel):
    title: str
    effective_date: Optional[str]
    governing_law: Optional[str]
    parties: list[Party]
    total_value: Optional[float]

out = readanydoc.extract(
    "contract.pdf",
    schema=Contract,
    instructions="Dates as YYYY-MM-DD. Amounts as numbers without currency signs.",
    array_extract=False,          # True for long documents with long lists: read in parts, merge the lists
    extract_model="claude-sonnet-5-5",
)

out.data              # a Contract instance
out.citations         # {"title": [...], "parties[0].name": [...], ...}
out.usage.cost_usd
out.warnings          # e.g. "required field 'governing_law' was not found in the document"

Argument	Meaning
`schema`	pydantic class or JSON Schema dict
`instructions`	system prompt for the extraction model
`citations`	default `True`
`array_extract`	process long documents in parts and merge their arrays
`mode`, `model`, `effort`, `model_options`	how pages are read (see Choosing models)
`extract_model`, `extract_effort`, `extract_options`	the model that fills the schema; default the reader's model
`pages`, `filename`, `timeout`	as in Parse
`settings`	any further Reducto extract setting

HTTP¶

curl -s -X POST http://localhost:8000/extract -H "content-type: application/json" -d '{
  "input": "reducto://...pdf",
  "instructions": {
    "schema": {"type": "object", "properties": {"total": {"type": "number"}}, "required": ["total"]},
    "system_prompt": "Be precise."
  },
  "settings": {"citations": {"enabled": true}, "array_extract": false}
}'

input can also be jobid://<parse job id> to reuse an earlier parse without reading the document again.

How ReadAnyDoc makes the output fit¶

Every reply is checked against your schema. Unambiguous spellings are converted ("$1,575.00" to 1575.0, "yes" to true); anything that still does not fit is re-asked a bounded number of times and then set to null, never guessed. Missing values are null, as in Reducto. Large schemas with many optional fields are fine: ReadAnyDoc does not rely on the provider's structured-output mode, which limits schema size.

Additions over Reducto¶

These are opt-in; a client that sends none of them behaves as on Reducto, plus the warnings list.

Field	What it does
`warnings` (response)	One line per required field that came back empty, at any depth
`instructions.examples`	Up to 5 example outputs, shown to the model as format examples. A returned value that equals an example and is not in the document is removed and reported
`settings.relevance_filter: true`	For documents of 4+ pages, one cheap call picks the pages that matter for your schema first. 80% cheaper with the same answers on an 86-page packet
`settings.request_pages: true`	The model may ask for more pages or search for a phrase (up to 3 rounds) before it answers. For cross-referenced contracts

Tips¶

Name fields the way the document names them, and describe units and formats in the field description.
Use Optional[...] for fields that may be absent; mark truly required fields as required so you get a warning when they are missing.
For long lists (statements, logs), set array_extract=True.