Lazy OCR, text layer, and PDF backend.

[GabHoo]

Question

Dear all,
I still do not have clear in my had the functioning of the pdf backend and what happens under the hood for a basic pdf conversion:
If I use the following pipeline:

pipeline_options = PdfPipelineOptions()
pipeline_options.do_ocr = False
pipeline_options.do_table_structure = False
pipeline_options.table_structure_options.do_cell_matching = False

doc_converter = DocumentConverter(
    format_options={
        InputFormat.PDF: PdfFormatOption(
            pipeline_options=pipeline_options, backend=PyPdfiumDocumentBackend
        )
    }
)

Even with this simple pipeline the conversion takes some time with a longer document. I am pretty sure there is a text layer already and if not I would like to skip the whole processing of document anyway since I will not have the resources do it (btw is there a way to check with Docling). I am forcingly disabling the process of OCR and table extraction. Why would than take more time than a normal pdf package to spit the text out?
I suppose that since Docling Document object contains much more information about the structure then just the text... is such rendering process that takes some time? What is it exactly happening under the hood (visual model/ rendering)?
I also tested the time for conversion to markdown and is pretty fast and accurate so I suppose that the converted document object contains all this structural information already. My last question is than: Is this rendering process using the existing text layer? What would it do with a pdf that does not have an embedded text layer already and OCR is disabled?

Thanks for your time and patience

[dosubot]

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

When you disable OCR and table extraction in Docling's PDF pipeline, the backend (like PyPdfiumDocumentBackend) still parses the PDF's text layer and extracts structural information—such as blocks, reading order, and layout—not just plain text. This structural parsing is more computationally intensive than what basic PDF text extraction tools do, which is why conversion can still take time, especially for longer or more complex documents. No visual model or image rendering is performed when OCR is off; only the text layer is used if present.

If the PDF has no embedded text layer and OCR is disabled, Docling will not extract any text content—the resulting document will be minimal or empty, and no fallback to image analysis or OCR occurs. The processing time you see is due to Docling's effort to build a rich internal document model, not due to rendering or OCR computation. Once this structure is built, conversions (like to markdown) are fast because all the heavy lifting is already done during the initial parse.
Source

To reply, just mention @dosu.

---

Share context across your team and agents. Try Dosu.