config.yaml

# PDF to Markdown Converter Configuration
# Optimized for Apple Silicon (M1/M2/M3/M4) with MLX acceleration

# Model Configuration
model:
  # Pipeline type:
  #   - vlm: Vision Language Model pipeline (fast, end-to-end, Apple Silicon optimized)
  #          Best for: Simple documents, speed priority, basic markdown output
  #          Does NOT support: OCR, table extraction, formula/code enrichment
  #
  #   - standard: Standard pipeline with full feature support (slower, more accurate)
  #          Best for: Complex documents, tables, formulas, scanned PDFs
  #          Supports: All features (OCR, tables, formulas, enrichments)
  pipeline_type: "standard"

  # VLM model type (only used when pipeline_type: vlm)
  # Options (Apple Silicon with MLX):
  #   - SMOLDOCLING_MLX: Fast, lightweight (256M params) - RECOMMENDED
  #   - GRANITEDOCLING_MLX: More accurate (258M params), slightly slower
  vlm_model: "SMOLDOCLING_MLX"

# Input/Output Paths
paths:
  # Input can be:
  #   - Single PDF file path
  #   - Directory containing PDFs (will process all PDFs)
  #   - URL to a PDF (e.g., https://arxiv.org/pdf/2408.09869)
  # Leave empty to use command-line argument
  input: ""

  # Output directory for converted files
  output: "./output"

  # Preserve input directory structure in output
  # Useful when processing nested directories
  preserve_structure: true

# Output Configuration
output:
  # Output format: markdown, json, html, doctags
  # Can be a single format or a list of formats for multiple outputs
  # Examples:
  #   format: "markdown"              # Single format
  #   format: ["markdown", "json"]    # Multiple formats
  format: ["markdown"]

  # Include images in output
  include_images: true

  # Image export mode:
  #   - referenced: Save images separately, reference in markdown
  #   - embedded: Embed images as base64 (larger files)
  image_mode: "embedded"

  # Image format: png, jpeg
  image_format: "png"

  # Image scale factor (1.0 = original size, 2.0 = double)
  image_scale: 1.0

# Processing Options
processing:
  # Enable OCR for scanned documents or images
  enable_ocr: false

  # OCR engine:
  #   - easyocr: Good balance of speed and accuracy
  #   - rapidocr: Fast, lightweight
  #   - tesseract: Most accurate, requires installation
  #   - ocrmac: macOS native OCR (fast, Apple Silicon optimized)
  #   - auto: Let Docling choose the best available
  ocr_engine: "auto"

  # Number of pages to process in parallel batch
  # Higher = faster but more memory usage
  # Recommended: 4 for 8GB RAM, 8 for 16GB+ RAM
  page_batch_size: 8

  # Maximum number of pages to process per document
  # 0 = unlimited
  max_pages: 0

  # Maximum file size in MB
  # 0 = unlimited
  max_file_size_mb: 0

  # Number of CPU threads for processing
  # 0 = auto (uses OMP_NUM_THREADS or defaults to 4)
  cpu_threads: 0

# Feature Toggles
# NOTE: Most features below only work with pipeline_type: standard
# VLM pipeline (faster) processes documents end-to-end and doesn't support these features
features:
  # Table structure recognition
  # Extracts tables with proper structure
  # Only works with: pipeline_type: standard
  table_structure: true

  # Use cell matching for tables
  # Maps recognized structure to PDF cells (more accurate)
  # Only works with: pipeline_type: standard
  table_cell_matching: true

  # TableFormer mode: FAST or ACCURATE
  # ACCURATE is slower but better for complex tables
  # Only works with: pipeline_type: standard
  table_mode: "ACCURATE"

  # Code block enrichment
  # Detects programming language in code blocks
  # Only works with: pipeline_type: standard
  code_enrichment: true

  # Formula enrichment
  # Extracts mathematical formulas as LaTeX
  # Only works with: pipeline_type: standard
  formula_enrichment: true

  # Picture classification
  # Classifies image types (chart, diagram, photo, etc.)
  # Only works with: pipeline_type: standard
  picture_classification: true

  # Picture description
  # Generates descriptions for images using VLM
  # Only works with: pipeline_type: standard
  # NOTE: Currently has issues with locally cached models
  picture_description: false

# Advanced Options
advanced:
  # Enable remote services (API calls to external services)
  # Set to true if using remote OCR or VLM services
  enable_remote_services: false

  # Custom model artifacts path
  # Empty = use default cache (~/.cache/docling/models)
  artifacts_path: ""

  # Verbose logging
  verbose: false

  # Fail on first error or continue processing
  continue_on_error: true

# Batch Processing (when input is directory)
batch:
  # Process files concurrently
  # 1 = sequential, >1 = parallel processing
  # Recommended: 1-2 for Apple Silicon to avoid memory issues
  max_workers: 2

  # File patterns to include (glob patterns)
  include_patterns:
    - "*.pdf"
    - "*.PDF"

  # File patterns to exclude
  exclude_patterns:
    - ".*" # Hidden files
    - "*~" # Backup files

  # Recursive directory traversal
  recursive: true

# Export Options (for different output formats)
export:
  # Markdown options
  markdown:
    # Include page breaks
    page_breaks: true

    # Image size in markdown (empty for original)
    image_width: ""

  # JSON options
  json:
    # Pretty print JSON output
    indent: 2

    # Include metadata
    include_metadata: true

  # HTML options
  html:
    # Include CSS styling
    include_styles: true

    # CSS style (light or dark)
    theme: "light"