Release notes

v1.0.0

pd3f 1.0 is one Python package and one Docker image: LiteParse replaces Parsr, the former pd3f-core package is part of pd3f, and Redis, RQ and the OCR container are gone.

To upgrade from 0.6, wait until no job is queued or running and remove the old setup with its volumes, before you pull 1.0, which no longer has these Compose files:

docker compose -f docker-compose.yml -f docker-compose.prod.yml down -v   # in a checkout of the repository
docker compose down -v                                                    # with the docker-compose.yaml from this website

Then start 1.0 as described in Installation. Old jobs are not taken over: delete ./data/pd3f-data-uploads/, or pd3f-data-uploads, pd3f-data-to-ocr and pd3f-data-cache next to the website’s Compose file.

  • the web app listens on port 8000 instead of 1616
  • the HTTP API moved to /jobs: POST /jobs instead of POST /, GET /jobs/<id> instead of GET /update/<id>, the outputs under /jobs/<id>/ instead of /files/, and DELETE /jobs/<id>; see Usage
  • the option fast became scorer=flair-fast; the default scorer none decides by rules and the document’s own spelling, without a language model
  • what experimental did is the default: headers and footers once at the start and the end, footnotes at the end, paragraphs cut by a page break joined where the text goes on; headers_footers, footnotes_in_place and page_breaks change it
  • tables always come as CSV; tables, parsr_config and parsr_adjust_cleaner_config are gone, check_ocr became ocr (auto, force, skip)
  • KEEP_RESULTS_HOURS became PD3F_JOB_TTL; RQ_REDIS_URL, REDIS_URL, WEB_CONCURRENCY, JOB_TIMEOUT, SENTRY_URL, DEMO, PRIVACY_URLS, MAX_UPLOAD_DISPLAY, POLL_INTERVAL, the form field in_browser, the RQ dashboard and the flask retry-failed and delete-failed commands are gone
  • the PyPI package pd3f carried pd3f-core up to 0.5.0; from 1.0.0 on it is this package: pd3f.extract returns a result for pd3f.output.text and pd3f.output.tables instead of the text and the tables
  • each line end, page break and choice between copies of a running line is logged in decisions.json, and the web app marks the uncertain ones
  • a command line tool, pd3f, and Markdown output with headings, lists and tables
  • OCR runs only on the pages that need it; scans are straightened and turned upright first
  • running headers and footers are recognized despite OCR’s misreadings of their copies
  • dehyphen 0.6 runs Flair’s language models with numpy, so the image has no PyTorch and is about 700 MB
  • the image runs natively on arm64, such as Apple Silicon
  • a new web app: several uploads at once, the text next to the PDF, light and dark mode, operable by keyboard

v0.6.x

To upgrade from 0.5, wait until the RQ dashboard shows no queued or running jobs, then stop everything and remove the volumes, because RQ 2 can’t read the results of the old jobs:

docker compose -f docker-compose.yml -f docker-compose.prod.yml down -v
git pull
docker compose -f docker-compose.yml -f docker-compose.prod.yml pull
./prod.sh -d

down -v also removes the cache volume, so the first job downloads the language models again, and the OCR files of old jobs in the to_ocr volume. Files of jobs from before the upgrade are not deleted automatically, remove them from ./data/pd3f-data-uploads/.

If you run pd3f with the docker-compose.yaml from this website, stop it with docker compose down -v, download the file again and start it with docker compose pull && docker compose up; files of old jobs stay in pd3f-data-uploads and pd3f-data-to-ocr. Own Compose files need RQ_REDIS_URL (e.g. redis://redis:6379) on web and worker, and the worker has to run rq worker --with-scheduler.

  • update to Python 3.13, Flask 3, RQ 2, Redis 8 and OCRmyPDF 17
  • the worker runs the RQ scheduler, so jobs and their files are deleted automatically KEEP_RESULTS_HOURS after they finished or failed; rq-scheduler and the rqscheduler command are gone
  • the Docker image uses the CPU build of PyTorch and is a lot smaller
  • pd3f-core 0.5 and dehyphen 0.4 no longer depend on Flair, the language models are loaded with a restricted unpickler
  • the RQ dashboard runs from the pd3f/pd3f image
  • the pd3f/pd3f and pd3f/pd3f-ocr images are published for amd64 and arm64
  • set REDIS_URL to use an external Redis
  • fix: the OCR worker did not start with current Docker versions, so every job waited forever
  • fix: flags such as fast=false were read as true
  • fix: example_api.py waited forever on failed jobs
  • fix: PDFs ending in .PDF waited forever for OCR
  • fix: the result page kept polling and showed no download buttons for jobs without tables, and did not show a single table
  • fix: text from the PDF was inserted into the result page as HTML
  • fix: the web app could cut off the first lines of a job log
  • fix: after a failed OCR, the result page showed an error on every poll once the files were deleted
  • fix: invalid JSON in parsr_config or parsr_adjust_cleaner_config caused an internal server error instead of status 400
  • ./prod.sh publishes the web app only on 127.0.0.1, like the dashboard; for access from other machines, put a reverse proxy in front, see In production with nginx
  • ./dev.sh publishes Flask’s development server only on 127.0.0.1
  • docker-compose.yml uses the published pd3f/pd3f-ocr image
  • table CSVs no longer contain an index column

v0.5.x

v0.4.x

  • improve Docker-Compose setup
  • add Italian support
  • update to Parsr v1.2.2, OCRmyPdf v11.6.2
  • make RQ-Dashboard v0.6.3 work by pinning RQ to v1.5.2
  • add optional timeout for rq jobs to mark started jobs that wont finish as failed

v0.3.x

  • first public release
Previous

Last updated on Oct 27, 2026