pdf-layout-qa

Compare a source PDF with a translation (or reprint) and mark layout mismatches: field alignment, lines, colors, fonts, sizes.

Layout extraction uses the Huridocs Docker image when it is up. Without it, the pipeline uses JSON next to each PDF (samples/ includes fixtures).

Run with Docker

First image build downloads PyTorch — expect several minutes.

python samples/generate_samples.py   # if samples/*.pdf are missing
docker compose up --build

Annotated PDFs land in Result/ on the host.

Skip Huridocs and use sidecar JSON only:

docker compose run --rm -e LAYOUT_SERVICE_URL= app python -m pdfqa

Run locally

Linux: sudo apt-get install poppler-utils

python -m venv .venv
# Windows: .venv\Scripts\activate
pip install -r requirements.txt
pip install -e .
python samples/generate_samples.py
python -m pdfqa

Optional layout service:

docker run --rm --name pdf-layout -p 5060:5060 --entrypoint ./start.sh huridocs/pdf-document-layout-analysis:v0.0.21

Override the URL: LAYOUT_SERVICE_URL=http://127.0.0.1:5060

API

from pdfqa.pipeline import compare_pdfs
from pdfqa.cli import configure_and_run

compare_pdfs("samples/form_en.pdf", "samples/form_es.pdf")
configure_and_run(pdf1="a.pdf", pdf2="b.pdf", color_threshold=80)

Thresholds: src/pdfqa/config.py

Layout

src/pdfqa/
  pipeline.py        # orchestrator
  cli.py             # config overrides
  layout/            # JSON sort, merge, semantic field match
  visual/            # lines, colors, fonts, crops, annotations
  util/
samples/             # synthetic EN/ES forms

Outputs: Result/doc1_annotated.pdf, Result/doc2_annotated.pdf

Do not commit third-party insurance forms. With all gratitute toward my friend Tony, whom I was too young to understand