Usage

Web app

Open http://localhost:8000 and upload one or more PDFs. The first job, or the one picked from the list, shows its steps while it runs, then the text next to the PDF. The uncertain decisions are marked; each mark can be focused and shows the options. Copy the text, or download it as text or Markdown, with the decision log, the tables as CSV and the OCRed PDF.

The browser remembers its own jobs. The first job with a Flair scorer in a language downloads that scorer’s two language models; flair-fast shares one pair for German, French and Italian.

Command line

pd3f letter.pdf > letter.txt
pd3f letter.pdf --lang en --headers-footers drop > letter.txt
pd3f letter.pdf --scorer flair --out letter/   # text.txt, text.md, decisions.json; tables/*.csv and ocr.pdf if any

The main options:

• --lang: de (default), en, es, fr or it; another language has no Flair scorer and needs --ocr-lang for OCR
• --scorer: what decides the cases the rules leave open, none (default), flair or flair-fast, see How it works
• --headers-footers: separate (default: one copy at the start and the end of the text), drop or keep in place, which leaves paragraphs cut by a page break apart
• --footnotes-in-place: keep footnotes at the end of their page instead of the end of the text; a paragraph cut by a page break then stays apart there
• --ocr: auto (default: pages without usable text), force or skip
• --out: a new or empty directory for the outputs instead of the text on stdout

Python

from pd3f import extract, output

result = extract("letter.pdf", lang="de", scorer="none")
print(output.text(result))

HTTP API

import time

import requests

with open("letter.pdf", "rb") as f:
    job = requests.post("http://localhost:8000/jobs", files={"pdf": f}, data={"lang": "de"}).json()

while (status := requests.get(f"http://localhost:8000{job['url']}").json())["state"] in ("queued", "running"):
    time.sleep(1)
if status["state"] == "failed":
    raise SystemExit(status["error"])
print(requests.get(f"http://localhost:8000{job['url']}/text.txt").text)

POST /jobs takes the PDF as pdf and the options of the command line as form fields: lang (one of the five), scorer, headers_footers, ocr, ocr_lang, and footnotes_in_place and page_breaks, which switch on when present. The outputs are files of the job: text.txt, text.md, decisions.json, ocr.pdf, tables/<n>.csv. DELETE /jobs/<id> deletes a job and its files. The README describes the status and the errors.

Housekeeping

The web app keeps each job in a directory under /data/jobs in the container (PD3F_JOBS) and the scorers’ models under /data/flair. Jobs older than PD3F_JOB_TTL hours (24) are deleted when a new job is created. PD3F_WORKERS (2) sets how many PDFs are extracted at the same time. Jobs that were queued or running when the container stopped are marked failed at the next start; upload those PDFs again.