PDF Pipeline Options

Every prior lesson already had options, just defaults

DocumentConverter() with no arguments, used in every beginner lesson, isn't option-free, it's running with PdfPipelineOptions()'s defaults: OCR on, table structure recognition on. Those defaults are why quarterly_report.pdf's table showed up correctly in Lesson 3 without any configuration from you. This lesson makes that configuration explicit, and shows what turning pieces of it off actually costs and buys.

Wiring options in through format_options

from docling.datamodel.base_models import InputFormat
from docling.datamodel.pipeline_options import PdfPipelineOptions
from docling.document_converter import DocumentConverter, PdfFormatOption

options = PdfPipelineOptions()
options.do_ocr = False
options.do_table_structure = False

converter = DocumentConverter(
    format_options={InputFormat.PDF: PdfFormatOption(pipeline_options=options)}
)

format_options is a dict keyed by InputFormat, this is also how you'd hand different options to PDF versus, say, a future format that takes its own FormatOption subclass. PdfFormatOption wraps a PdfPipelineOptions instance specifically for the PDF pipeline.

What do_ocr=False and do_table_structure=False change

Turning both off skips the two most expensive per-page models. On quarterly_report.pdf, which already has a real text layer (it's a generated, not scanned, PDF), do_ocr=False loses nothing: the text layer is read directly either way, OCR only matters when there's no text layer to read (Lesson 9). do_table_structure=False does change something: a table item is still emitted from the PDF's own cell grid, but without TableFormer reconciling merged cells and header rows, so structure fidelity on a genuinely complex table would drop even though a table object still exists.