Document Parsing

DocLD supports parsing multiple document formats with intelligent layout detection, OCR, and semantic chunking.

Supported Formats

Processing Pipeline

Upload → Parse → OCR → Chunk → Vectorize → Index

1. Upload — File is uploaded to secure storage (Cloudflare R2)
2. Parse — Content extracted based on file type
3. OCR — VLM-based OCR handles images and scanned pages
4. Chunk — Semantic and layout-aware chunking
5. Vectorize — Chunk text is sent to Pinecone; embeddings are generated server-side (llama-text-embed-v2)
6. Index — Records stored in Pinecone for semantic search

OCR Capabilities

Language Support

DocLD supports 50+ languages for OCR:

• European: English, Spanish, French, German, Italian, Portuguese, Dutch
• Asian: Chinese (Simplified/Traditional), Japanese, Korean, Thai, Vietnamese
• Middle Eastern: Arabic, Hebrew, Persian
• Indic: Hindi, Bengali, Tamil
• Auto-detection when language not specified

OCR Features

OCR Confidence

Results include confidence scores:

{
  "text": "Extracted text content",
  "confidence": 0.95,
  "language": "en",
  "bounding_boxes": [
    {
      "text": "Invoice #12345",
      "x": 100,
      "y": 50,
      "width": 150,
      "height": 25,
      "confidence": 0.98
    }
  ]
}

Agentic OCR

Enable AI-enhanced parsing for complex documents:

{
  "parsing_config": {
    "agentic": true
  }
}

Agentic mode provides:

• Better table reconstruction
• Figure and chart analysis
• Layout understanding
• Complex form parsing
• Higher accuracy (slower processing)

Chunking Strategies

Semantic Chunking (Default)

Splits by meaning and natural boundaries:

• Respects paragraph breaks
• Keeps related content together
• Preserves context across chunks

Fixed Size Chunking

Splits by character count:

{
  "chunking": {
    "strategy": "fixed",
    "max_chunk_size": 1000,
    "overlap": 100
  }
}

Page-Based Chunking

One chunk per page:

{
  "chunking": {
    "strategy": "page"
  }
}

Configuration Options

Parsing Config

API parse endpoint (POST /api/parse): accepts config with formatting.table_output_format and chunking (strategy, max_chunk_size, overlap). See the Parse API for the exact request shape.

Full document processing (e.g. document ingestion, workflows, or processing via the dashboard): uses a richer parsing config that can include OCR, agentic mode, and metadata options:

{
  "parsing_config": {
    "chunking": {
      "strategy": "semantic",
      "max_chunk_size": 1000,
      "overlap": 100
    },
    "ocr": {
      "enabled": true,
      "language": "auto"
    },
    "agentic": false,
    "include_metadata": true
  }
}

Parsing Methods

Synchronous

Parse immediately and get results:

curl -X POST "/api/parse" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -F "file=@document.pdf"

Asynchronous

Queue for background processing:

curl -X POST "/api/parse/async" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -F "file=@document.pdf" \
  -F "webhook_url=https://your-server.com/webhook"

Output Format

Parse results include structured content:

{
  "job_id": "uuid",
  "duration": 2.5,
  "usage": {
    "num_pages": 5,
    "credits": 7.5
  },
  "result": {
    "type": "full",
    "chunks": [
      {
        "content": "# Invoice\n\nInvoice Number: INV-001...",
        "page": 1,
        "metadata": {
          "type": "text",
          "confidence": 0.98
        }
      }
    ]
  }
}

Credit Usage

CLI Support

Parse documents from the command line:

# Parse a single file
docld parse document.pdf
 
# Parse with agentic mode
docld parse document.pdf --agentic
 
# Parse an entire folder
docld parse ./documents -o ./parsed

API Reference

See the Parse API for complete endpoint documentation.