Can someone help me understand, the use/need for an OCR engine when using something like granite-docling?

[icsy7867]

Apologies for the stupid question. But I am trying to understand what exactly is needed, and the benefits of using, or not using OCR.

From my understand, something like Granite-docling should replace the need for a third party OCR tool/engine shouldnt it? As the model should understand an image/document on it's own.

For simplicity, when using the docling UI:

What exactly changes when I check or uncheck "Enable OCR"? If using a document reading capable model, is OCR necessary? Is it beneficial to always use it?

Thank you!

[dosubot[bot]]

Great question—here’s how it works:

When you check "Enable OCR" in the Docling UI (with the Standard pipeline), Docling runs an OCR engine (like EasyOCR, Tesseract, or RapidOCR) to extract text from image-based or scanned regions of your PDF. This is essential for scanned documents or PDFs without a selectable text layer—otherwise, Docling won’t extract any text from those pages, and your output will be empty or incomplete. If you uncheck "Enable OCR," Docling only uses the embedded text layer (if present) and ignores any text that’s part of an image, so scanned or image-only PDFs won’t be processed for text at all. OCR can slow things down, but it’s necessary for recovering text from non-digital sources and visually complex layouts. You can also fine-tune when OCR is applied using options like bitmap area thresholds or force full-page OCR, but there’s no built-in way to auto-toggle OCR only when needed [source].

Granite-docling is a Vision-Language Model (VLM) and is used in the "vlm" pipeline, which is separate from the standard pipeline. The "Enable OCR" toggle in the UI does not affect the VLM pipeline—granite-docling processes images and documents directly using its own vision-language capabilities. So, if you’re using the VLM pipeline, the OCR setting is ignored [source].

In summary:

• Use OCR for scanned/image-based PDFs in the Standard pipeline.
• For digital PDFs with a proper text layer, you can disable OCR for faster processing.
• Granite-docling (VLM pipeline) doesn’t use the OCR toggle; it handles images itself.

Here’s a simple diagram of the Standard pipeline flow:

graph LR
    A["PDF Input"] --> B["Preprocessing"]
    B --> C{"Enable OCR?"}
    C -- Yes --> D["OCR Engine (extract text from images)"]
    C -- No --> E["Use embedded text layer only"]
    D --> F["Layout/Table/Assemble"]
    E --> F
    F --> G["Output (JSON, Markdown, etc.)"]

Loading

Let me know if you want more details on pipeline selection or OCR engine options!

To reply, just mention @dosu.

---

Share context across your team and agents. Try Dosu.