Skip to content

0.3.0

Choose a tag to compare

@arpitpatel25 arpitpatel25 released this 18 Nov 19:11
· 19 commits to main since this release
587d451

Every extraction call is enforced by the API, whatever the schema's size, and reads one cached copy of the
document; citations point where a person checking a value would look; and a cheaper one-stage mode.

  • Enforced output at any size. A schema the API refuses as one structured-output call is asked in one fixed answer
    format ({"values": [{"code", "item", "value"}]}, codes from your schema's own fields), the same in every call,
    and turned back into your exact shape. No call is ever made just to fix the shape.
  • The document is cached within a run: after the first group of fields, every group reads it at a tenth of the
    price. Nothing is reused between runs.
  • Page parts are sized in tokens: part_tokens (default 30,000, capped by the model's own limit; extract(..., part_tokens=...), settings.part_tokens, MCP part_tokens). Measured on a 43-page 10-Q: 30,000 kept the
    accuracy of small parts at lower cost; one 43-page part lost a third of the values.
  • Model limits (context, output, price jump, prices) come from a snapshot of LiteLLM's public model list, about
    4,300 models, or an OpenAI-compatible server's /models; your own file in READANYDOC_MODEL_LIMITS overrides
    them. An unknown model is an error naming part_tokens, never a guess. The output limit is the model's own.
  • One log line per run with every setting chosen (readanydoc logger, extract settings: {...}), and
    READANYDOC_TRACE=<file.jsonl>: one line per model call (tokens, cache reads and writes, cost, seconds, why the
    call was made, the reply) and per decision.
  • A value stated once for a whole table or page (a unit, "in thousands", a period in a heading) is written on every
    item it covers; of two copies of the same list item, the less complete one is dropped; names are copied in full.
  • Citations: the table the model read each row from is kept across page parts; worked-out values (a period from a
    column heading, a scale from "in millions", a code from a name) are boxed on the printed words they come from;
    a category chosen from your schema's list is never boxed on the same word elsewhere; a number is matched only as
    a whole number; a short value printed in several places is not boxed at a guess. On a 10-Q, numbers boxed on the
    right row: 94% (Reducto 81%); category boxes on the cover page: 128 to 0.
  • extract(..., parse=False) (MCP parse_first=false): one stage, off by default. The model reads the file itself
    (PDF, scans, images, Office files) and citations come from the PDF's own words or local OCR, with no model. About
    half the cost and 2.5 times faster on a 43-page 10-Q, but less accurate (102-110 values against 99-129) and with
    3-4 times fewer citations.
  • A value the document does not give stays null, also for a required field: never an invented value.

Full research record: docs/research/extraction.md.