obliq-ocr

Pulls the text out of PDFs and scans, one page at a time. Pages that already have digital text are read straight from the PDF in milliseconds; scanned pages go through OCR (RapidOCR, PP-OCRv4 on ONNX Runtime). A file that mixes both is handled without any settings.

It was built at OBLIQ for the documents CA firms get from their clients: tax invoices, bills, challans, e-way bills and compiled PDFs. It's for Python developers who need that text as plain strings or JSON, and its output is meant to be checked by a person before it's used.

Install

pip install obliq-ocr

• Python 3.10, 3.11 or 3.12. The OCR engine doesn't install on 3.13 yet.
• Linux and Windows are tested. macOS should work but isn't tested.
• Runs on CPU. For an NVIDIA GPU, see GPU.

The first release isn't on PyPI yet. Until it is, install from GitHub:

pip install "git+https://github.com/OBLIQ-in/obliq-OCR.git"

Quickstart

from obliq_ocr import parse_document, save_result

result = parse_document("invoice.pdf")  # a PDF or an image

for page in result.pages:
    print(page.page, page.method, page.confidence)

print(result.full_text())
save_result(result, out_dir="outputs")  # outputs/invoice.json and outputs/invoice.txt

parse_document prints one progress line per page; pass verbose=False to turn that off.

Supported files

Sample output

save_result writes this JSON (the same as result.to_dict()). This is the real output for a two-page test PDF where page 1 is digital and page 2 is a scan:

{
  "source": "invoice.pdf",
  "device": "CPU",
  "pages": [
    {
      "page": 1,
      "method": "text_layer",
      "reason": "97 chars, image cover 0%",
      "text": "Mixed document digital page.\nThis page contains native text layer elements.\nIt should be extracted via text layer.",
      "n_chars": 114,
      "confidence": null,
      "seconds": 0.018
    },
    {
      "page": 2,
      "method": "ocr",
      "reason": "only 0 text chars",
      "text": "Mixed document scanned page.",
      "n_chars": 28,
      "confidence": 1.0,
      "seconds": 0.516
    }
  ]
}

method is text_layer, ocr or blank. confidence is the average OCR score from 0 to 1, and null for pages that weren't OCR'd. Every field and error is explained in the architecture doc.

Command line

obliq-ocr invoice.pdf                                   # writes outputs/invoice.json and .txt
obliq-ocr scan.pdf --output-dir results --device cpu    # choose folder and device
obliq-ocr scan.pdf --quiet                              # no per-page progress
python -m obliq_ocr invoice.pdf                         # same thing, as a module

GPU

pip uninstall -y onnxruntime
pip install "obliq-ocr[gpu]"

The CPU and GPU builds of ONNX Runtime can't be installed together, which is why the first command is needed. obliq-ocr uses CUDA when it's available; set OBLIQ_OCR_DEVICE=cpu or cuda, or pass force_device="cpu" or "cuda", to choose.

Limitations

The model card has the measurements. In short:

• Scanned pages can't produce ₹ or Hindi and other Indian scripts. Digital PDFs keep them.
• Handwriting isn't supported.
• Low-resolution or skewed scans can change digits and drop whole lines.
• A high confidence doesn't mean the text is right: check key values.
• The output has no coordinates yet (#71).

Documentation

• Architecture and output schema: how pages are routed, every output field, errors
• Model card: models, versions, measured accuracy, limitations
• Benchmark: speed and memory against other OCR engines
• Changelog
• Release guide, for maintainers
• Security policy: how to report a vulnerability privately

Contributing

git clone https://github.com/OBLIQ-in/obliq-OCR.git
cd obliq-OCR
python -m venv .venv            # Python 3.10 to 3.12
source .venv/bin/activate       # Windows: .venv\Scripts\activate
pip install -e ".[dev,test]"
pre-commit install
python tests/scripts/generate_fixtures.py   # synthetic test PDFs
pytest

• Branch off main and open a PR. PR titles follow Conventional Commits (docs: …, fix: …, feat: …). If the PR has a single commit, its message must follow the same format.
• CI runs ruff, mypy, the tests on Python 3.10 to 3.12, an accuracy gate and a link check.
• Never commit real client documents or anything with personal data. Test files must be synthetic or fully anonymised.

Questions and bugs go to GitHub issues.

License

MIT