Docling contains the following pipelines, each with specific purposes and selection criteria:

Production Pipelines#

1. StandardPdfPipeline

   • Purpose: Multi-threaded PDF processing with layout detection, OCR, and table structure extraction.
   • Auto-Selected For: PDF, IMAGE, METS_GBS formats.
   • Manual Only: No (default for relevant formats).
   • Strengths: Always provides bounding boxes, well-tested, faster due to threading.
   • Limitations: Traditional OCR may struggle with complex layouts.
   • Scanned Documents: This is the default pipeline for scanned documents, using OCR (EasyOCR by default, with options for Tesseract or RapidOCR).

2. SimplePipeline

   • Purpose: Direct conversion for structured formats.
   • Auto-Selected For: DOCX, PPTX, XLSX, HTML, Markdown (.md), plain-text (.txt, .text), Quarto (.qmd), R Markdown (.Rmd), AsciiDoc, LaTeX, CSV, JSON_DOCLING, VTT, XML variants.
   • Manual Only: No (default for these formats).

3. VlmPipeline

   • Purpose: Vision-Language Model document understanding.
   • Auto-Selected For: None.
   • Manual Only: Yes (must be selected with --pipeline vlm).
   • Strengths: Better semantic understanding of complex layouts, tables, and figures.
   • Limitations: More computationally expensive; bounding boxes only available with DocTags presets.
   • Scanned Documents: Can be used for scanned documents when semantic understanding is critical, but not auto-selected.

4. ExtractionVlmPipeline

   • Purpose: Schema-based structured data extraction using VLM models (NuExtract, Granite Vision).
   • Auto-Selected For: None.
   • Manual Only: Yes.
   • Model Selection: Configurable via VlmExtractionPipelineOptions (default: NuExtract-2B, alternative: Granite Vision 4.1).

5. AsrPipeline

   • Purpose: Speech-to-text transcription using Whisper models.
   • Auto-Selected For: Audio files.
   • Manual Only: No (default for audio).

Experimental Pipeline#

• ThreadedLayoutVlmPipeline
  • Purpose: Hybrid layout model detection plus VLM processing with spatial context.
  • Status: Experimental, not production ready.

Deprecated/Alias Pipelines#

• LegacyStandardPdfPipeline: Deprecated, replaced by StandardPdfPipeline.
• ThreadedStandardPdfPipeline: Alias for StandardPdfPipeline.

Scanned Documents Clarification#

• StandardPdfPipeline with OCR is the default and only auto-selected pipeline for scanned documents. It uses OCR to extract text and layout information.
• VlmPipeline can be manually selected for scanned documents when advanced semantic understanding is needed, but is not auto-selected.
• The previous mention of "another pipeline" for scanned documents referred to StandardPdfPipeline with OCR enabled, not a separate pipeline.

Pipeline Comparison for Scanned Documents#

No other pipelines exist in Docling for scanned documents.

For more details, see the Docling repository pipelines.