Is OCR available for embedded images in pptx and docx?

[chris-prezi]

Is OCR only available for embedded images in pdf, or can we extract text from images in pptx and docx as well?

[dosubot[bot]]

OCR is currently only available for PDF documents—it's not supported for embedded images in PPTX or DOCX files.

The reason is the different pipelines used:

• PDFs use StandardPdfPipeline which includes built-in OCR support (configurable via do_ocr, force_ocr, ocr_engine, etc.)
• PPTX and DOCX use SimplePipeline which extracts text, images, and tables directly from the document structure, but doesn't apply OCR to embedded images

If you need to extract text from images embedded in PowerPoint or Word documents, your options would be:

1. Convert the documents to PDF first, then process with OCR enabled
2. Use a separate OCR tool on the extracted images

To reply, just mention @dosu.

---

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

   

[dosubot[bot]]

Yes, that multi-pass approach would work. You'd:

1. Convert PPTX/DOCX with Docling to get text + extract images via doc.pictures
2. Run OCR separately on those extracted PIL images (using an external OCR library like EasyOCR, Tesseract, etc.)

That said, a simpler alternative might be to convert the PPTX/DOCX to PDF first (e.g., via LibreOffice), then process the PDF with Docling's StandardPdfPipeline which has built-in OCR support. This gives you a single-pass workflow with native OCR integration.

To reply, just mention @dosu.

---

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

   

[chris-prezi]

@dosu - I had thought we might lose some structural context from pptx and docx documents if they were converted to PDF. Would we lose anything? I assume that docling understands the structure of pptx and docx, so can accurately convey that in the extracted markdown. Is any of that structural context lost by first converting to PDF?

[dosubot[bot]]

You're right to be concerned—there are trade-offs both ways:

What you'd likely lose by converting to PDF first:

• DOCX headers/footers are parsed as structured GroupItem objects with content layer attribution (FURNITURE vs BODY). Converting to PDF flattens this distinction
• PPTX extracts text with paragraph styles, list indentation, and master styles directly from the PowerPoint object model—PDF conversion would require layout reconstruction
• Slide notes from PPTX are extracted directly but may not survive PDF conversion depending on your conversion tool

What you'd gain by converting to PDF first:

• DOCX files don't provide page numbers or bounding box metadata—if you need spatial provenance for chunking/retrieval, PDF gives you that
• OCR support for embedded images (as discussed)

So it's a trade-off: direct parsing preserves more semantic structure, but PDF gives you spatial metadata + OCR. Depending on your use case, you might even consider a hybrid approach—direct parsing for structure, with selective OCR on extracted images.

To reply, just mention @dosu.

---

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

   

[chris-prezi]

@dosu - we're currently migrating to docling-serve, from docling via AWS Sagemaker. Can we obtain all the image information we would need to adopt a multi-pass solution (text, then image OCR) from docling-serve's responses?

[dosubot[bot]]

Yes! Docling-serve supports image extraction with a few options for how images are delivered. The key parameter is image_export_mode:

• embedded (default) - images are base64-encoded inline in the response
• referenced - images are saved as separate files (requires target_type: "zip")
• placeholder - images replaced with placeholder text (no image data)

For a multi-pass OCR workflow, I'd recommend using referenced mode with zip output:

{
  "options": {
    "image_export_mode": "referenced",
    "target_type": "zip",
    "include_images": true,
    "images_scale": 2.0
  }
}

This gives you the extracted images as separate files in the zip archive, with PictureItem objects containing URIs pointing to those image files. You can then run your separate OCR pass on those extracted images.

To reply, just mention @dosu.

---

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