Guides

Convert Arabic PDFs to text in Python

Extract Arabic text from scanned or image-only PDFs, with tashkeel and tables, using the open-source Nuqta OCR library.

1. Install

pip install "nuqta-ocr[vllm]"   # Linux + NVIDIA GPU, fastest
pip install "nuqta-ocr[fast]"   # transformers with fast kernels
pip install nuqta-ocr           # transformers only

vLLM is the fastest backend (about 3 seconds per page for the 0.8B on a 24 GB GPU). Without a GPU, see running offline on CPU.

2. Read the PDF

from nuqta import OCR

if __name__ == "__main__":
    ocr = OCR("nuqta-4b")                                 # or "nuqta-0.8b"
    result = ocr.read("contract.pdf", mode="markdown")    # every page, rendered at 200 dpi
    for i, page in enumerate(result.pages, 1):
        print(f"--- page {i} ---")
        print(page.text)
    open("contract.md", "w", encoding="utf-8").write(result.text)

mode="markdown" keeps headings, lists and tables (as HTML). Use the default as-written mode for plain text, or bare to remove all tashkeel.

3. Many files at once

results = ocr.read_many(["a.pdf", "b.pdf", "scan.jpg"], mode="markdown")   # batched on the GPU

Command line

nuqta read contract.pdf --mode markdown > contract.md

Why not pdftotext?

Text extraction tools such as pdftotext or PyMuPDF only work when the PDF already contains a text layer. Scanned PDFs are images, and many Arabic PDFs have a broken text layer with letters in the wrong order or in presentation forms. OCR reads what is printed.

Questions

How do I extract Arabic text from a scanned PDF?

Use an OCR model, because a scanned PDF contains images, not text. With the open-source nuqta-ocr library: pip install nuqta-ocr, then OCR('nuqta-4b').read('file.pdf').text.

Why is Arabic text from my PDF reversed or broken?

The PDF's text layer stores Arabic in visual order or as presentation-form glyphs. OCR the rendered page instead of copying the text layer.