How to Extract Scanned Arabic Text?

[abedkhooli]

Question

I am trying to convert a test document to Markdown. The doc has both English and Arabic (live text and clean text as image for each). Text extraction works in the standalone RapidOCR library (using params:

       "Rec.engine_type": EngineType.ONNXRUNTIME, 
        "Rec.lang_type": LangRec.ARABIC, # ARABIC uses python-bidi
        "Rec.model_type": ModelType.MOBILE,
        "Rec.ocr_version": OCRVersion.PPOCRV5,

with python-bidi installed.
I tried from Docling (passed the .onnx path in RapidOcrOptions and patched the model -> PatchedRapidOcrModel(RapidOcrModel)) but no success.
Does Docling currently support languages other than English and Chinese?
If yes, how do I achieve that?
Here's a sample file for testing:
docling_test_ak.pdf

[dosubot]

Hi @abedkhooli! I'm Dosu and I’m helping the docling team.

Unfortunately, Docling's RapidOCR integration does not currently expose the advanced parameters you need for Arabic OCR (like LangRec.ARABIC, ModelType.MOBILE, OCRVersion.PPOCRV5). The lang parameter in RapidOcrOptions is explicitly documented as non-functional, and Docling always loads Chinese PP-OCRv4 models by default regardless of language settings.

Known Arabic OCR issues: There are existing reports of Arabic text being reversed (RTL appears as LTR) and garbled output, suggesting this isn't well-supported yet.

Possible workarounds:

1. Try the rapidocr_params dictionary (most promising):

from docling.datamodel.pipeline_options import PdfPipelineOptions, RapidOcrOptions

ocr_options = RapidOcrOptions(
    backend="onnxruntime",
    det_model_path="/path/to/arabic_det.onnx",
    rec_model_path="/path/to/arabic_rec.onnx", 
    cls_model_path="/path/to/arabic_cls.onnx",
    rec_keys_path="/path/to/arabic_dict.txt",
    rapidocr_params={
        # Additional RapidOCR parameters - may allow some advanced config
    }
)

2. Specify custom Arabic model paths directly if you have the Arabic PP-OCRv5 ONNX models.

However, the core issue is that LangRec.ARABIC (which enables python-bidi for RTL handling) isn't configurable through Docling's interface. This may require subclassing RapidOcrModel as you mentioned, or contributing a feature request to expose these parameters.

Related discussions: #1421, #2927, #1727

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}

   

[abedkhooli]

@dosu RapidOCR is in version 3.6 now and there is no det model for Arabic. Providing path did not work and even hardcoding paths as in #2927 did not make a difference. I suggest passing this to a human developer.

[dosubot]

You're right that this is a fundamental limitation that requires developer attention. Multiple users have reported that RapidOCR completely fails with Arabic documents, and the language parameter is explicitly non-functional and reserved for future compatibility.

This issue is related to existing tickets:

• #1421 - Arabic language support problems (RTL text handling)
• #541 - PaddleOCR discussion, where users note it handles Arabic better than RapidOCR

The maintainers who typically handle OCR-related issues include @cau-git, @nikos-livathinos, and @dolfim-ibm. Adding a label or tagging them directly may help escalate this for proper Arabic OCR support.

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}

   

[abedkhooli]

@cau-git, @nikos-livathinos, and @dolfim-ibm FYI if there is a workaround.