Document Processing¶

Extract structured text and layout from documents (PDFs, images).

Processors¶

Docling¶

Best for: Advanced layout understanding, complex documents

docs2synth preprocess document.pdf --processor docling

Advanced layout analysis
Table and figure extraction
Multi-page support
Reading order detection

PaddleOCR¶

Best for: General OCR, multilingual documents, Asian languages

docs2synth preprocess document.png --processor paddleocr --lang en

80+ languages
GPU acceleration
Bounding boxes with confidence scores
Fast and accurate

PDFPlumber¶

Best for: Parsed PDFs with text layers (digital documents)

docs2synth preprocess document.pdf --processor pdfplumber

Very fast (no OCR needed)
Word-level bounding boxes
Table extraction
Only works with text-based PDFs

EasyOCR¶

Best for: Alternative OCR, 80+ languages

docs2synth preprocess document.png --processor easyocr

80+ languages
GPU acceleration
Easy setup
Good for scanned documents

Usage¶

# Single document
docs2synth preprocess document.png

# Directory
docs2synth preprocess ./documents/

# With options
docs2synth preprocess ./documents/ \
  --processor docling \
  --output-dir ./processed \
  --lang en \
  --device gpu

Output Format¶

JSON files with: - objects: Text objects with text, bbox, page, label - context: Full document text in reading order - reading_order_ids: Object sequence - process_metadata: Processor info, timestamp - document_metadata: Filename, dimensions

Example:

{
  "objects": {
    "obj_0": {
      "text": "Invoice Number: INV-2024-001",
      "bbox": [100, 200, 300, 220],
      "page": 0,
      "label": "text"
    }
  },
  "context": "Invoice Number: INV-2024-001\n...",
  "reading_order_ids": ["obj_0", "obj_1", ...],
  "process_metadata": {
    "processor_name": "docling",
    "timestamp": "2024-01-15T10:30:00Z"
  }
}

Processor Comparison¶

Processor	Speed	Languages	Tables	OCR	Best For
Docling	Medium	All	✅	✅	Complex layouts
PaddleOCR	Fast	80+	❌	✅	General OCR
PDFPlumber	Very Fast	All	✅	❌	Text PDFs
EasyOCR	Medium	80+	❌	✅	Alternative OCR

Next Steps¶

QA Generation - Generate QA pairs
Retriever Training - Train models
RAG Deployment - Build RAG systems

For complete workflow: Complete Workflow Guide

API Reference¶

See API Reference page.