Docling Config Generator

The Config Generator turns your selections into a copy-ready docling convert command and an equivalent Python snippet. This guide explains what every option does, how the generated flags map to the Python API, and how to run the result safely.

Work top to bottom; the preview refreshes as you change any control.

1. Choose the input format that matches your file. It only changes input handling for PDFs; other formats use their own backend.
2. Decide on OCR. Leave it on for scans and photos, and turn it off for digital PDFs to make conversion much faster.
3. Pick an OCR engine and mode if OCR is on. Start with auto and choose a specific engine only when you need it.
4. Select the table mode. Use accurate for complex tables and fast when speed matters more than precision.
5. Enable only the enrichment you need. Code, formulas and charts each add neural post-processing.
6. Choose the output format and accelerator. Markdown for reading, JSON for structure; pick a GPU only if you have one.
7. Copy, or download the script. Use Copy CLI, Copy Python or Download Script (.sh).
8. Run it and check the output. The converted file is written next to your source by default.

Pick the closest match. Only PDF and image inputs are affected by the OCR and table options; the rest flow through their format-specific backend automatically.

OCR (optical character recognition) converts text on images into machine-readable text. It is the single biggest factor in both accuracy and runtime, so enable it deliberately.

• Enable OCR when the document is a scan, a photo, or a PDF without a selectable text layer.
• Disable OCR for digital PDFs - it can make conversion several times faster (--no-ocr).
• default mode keeps existing text and only OCRs pages that are missing text.
• full_page mode OCRs every page and overwrites the detected text.

Tables are reconstructed from layout, not from raw text, so the mode you choose changes the result quality.

Enrichment adds neural post-processing to detect specialised content. Enable it only when the document actually contains it.

Choose what you want to do with the result; the preview also updates the export call in the Python snippet.

Docling runs on CPU, on NVIDIA GPUs (CUDA) and on Apple Silicon (MPS). auto selects CUDA, then MPS, then CPU.

• Use --device cuda on NVIDIA hardware for large batches.
• Use --device mps on Apple Silicon.
• On CPU, disable OCR and enrichment you do not need.

These are the flags the generator can emit, with the Python option each one maps to. For the complete CLI surface, see the CLI reference.

The generator builds this snippet for you. For PDFs it configures a PdfPipelineOptions object and registers it through format_options; for every other format a plain DocumentConverter() is enough.

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

pipeline_options = PdfPipelineOptions()
pipeline_options.do_ocr = True
pipeline_options.do_table_structure = True

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

result = converter.convert("report.pdf")
print(result.document.export_to_markdown())

• Swap export_to_markdown() for export_to_dict() to get lossless JSON.
• Set pipeline_options.do_ocr = False to disable OCR.
• Use convert_all() to process a batch of files.

Copy a preset and change the file name.

The generator checks combinations and warns you when a choice will not take effect. Always confirm unusual flags against the official CLI reference.

• Audio input selected while OCR is on (OCR does not apply to the ASR pipeline).
• OcrMac selected on a non-macOS platform.
• Nemotron OCR selected on CPU.
• OCR disabled but an OCR engine still selected.

The generator runs entirely in your browser, and the commands it produces run Docling locally.

• Disable OCR for digital PDFs; keep it only for scans.
• Use --table-mode fast when approximate tables are acceptable.
• Prefer a GPU for large batches, and tune --num-threads to your machine.
• Nothing is uploaded: your selections never leave the browser.