RTL languages & Text Rendering Issues in PDF Knowledge Base Imports

  • Polaris

    (@namrur)


    6 days, 9 hours ago

    Hello,

    We are experiencing a significant text-rendering issue within the Knowledge Base Manager when importing or processing PDF files that contain Right-to-Left (RTL) languages.

    Currently, the internal PDF processing/parsing library used by your plugin does not support RTL script properly. As a result, text extracted from uploaded PDFs is reversed (rendered completely backward and unreadable).

    For example, taking the English equivalent sentence “My dog has learned a new trick“, the text appears like this:

    “kcirt wen a denrael sah god yM”

    This makes the content entirely illegible, as all uploaded PDF content becomes unreadable. In addition, because the text is indexed in reverse order, searching the knowledge base is practically impossible, as you most type search keywords backward to get a matchs.

    To properly support RTL languages and character shaping, I would like to recommend updating the PDF handling library or utilizing a robust library like mPDF (which natively handles RTL text shaping, BiDi algorithms, and font embedding for right-to-left languages as well).

    Could you please look into this or advise if there is a workaround or a planned update to address RTL text extraction?

    Thank you!

Viewing 1 replies (of 1 total)
  • Plugin Support m4xw3ll

    (@m4xw3ll)

    6 days, 7 hours ago

    Hi @namrur,

    Thanks for the detailed report – confirmed. Text extraction currently returns RTL PDFs in visual order whenever the PDF’s producer doesn’t tag its right-to-left runs, which is why the stored text comes out reversed. (PDFs exported from Microsoft Word usually carry that tag and import correctly, which is why this doesn’t affect every file.)

    A fix is scheduled: extraction will detect Hebrew/Arabic content and restore logical character order before anything is indexed, so both readability and knowledge base search work as expected. It applies at import time, so you’ll want to re-import your existing RTL PDFs once the update lands.

    On mPDF – it’s a PDF generation library rather than a text extractor, so it can’t be swapped in here, but your underlying point stands and we’re fixing it at the extraction layer.

    In the meantime, the manual text entry and website/URL import paths handle RTL correctly – only PDF extraction is affected. If your PDF content also exists as web pages, importing those URLs will index properly today.

    I’ll follow up on this thread when the release ships.

Viewing 1 replies (of 1 total)

You must be logged in to reply to this topic.