Document Parser Skill

Purpose

Extracts text and structured data from documents using OCR technology, enabling automated processing of invoices, receipts, statements, and contracts.

Triggers

• •PDF or image document uploaded
• •Scanned invoice needs processing
• •Receipt needs expense categorization
• •Bank statement needs reconciliation

Capabilities

1. •OCR Text Extraction - Convert images to text
2. •PDF Parsing - Extract text from PDF documents
3. •Structured Data Extraction - Identify fields, tables, line items
4. •Document Classification - Classify document type
5. •Quality Assessment - Assess OCR confidence

Instructions

Step 1: Document Type Detection

Classify document:

• •Invoice - Vendor bill
• •Receipt - Proof of purchase
• •Bank Statement - Monthly statement
• •Contract - Legal agreement
• •Other - Unknown type

Use file name, content patterns, or LLM classification.

Step 2: OCR Processing

Google Document AI (Preferred)

const document_ai = require('@google-cloud/documentai');
const client = new document_ai.DocumentProcessorServiceClient();

const [result] = await client.processDocument({
  name: processor_name,
  rawDocument: {
    content: file_buffer.toString('base64'),
    mimeType: 'application/pdf',
  },
});

const extracted_text = result.document.text;
const entities = result.document.entities;  // Pre-extracted fields

Tesseract OCR (Fallback)

tesseract invoice.png output -l eng

Step 3: Field Extraction

For Invoices, extract:

• •Vendor name and address
• •Invoice number and date
• •Due date
• •Line items (description, quantity, price)
• •Subtotal, tax, total

For Receipts, extract:

• •Merchant name
• •Date and time
• •Items purchased
• •Total amount

For Bank Statements, extract:

• •Account number
• •Statement period
• •Transaction list (date, description, amount)
• •Beginning and ending balance

Step 4: Structured Output

Return JSON with confidence scores:

{
  "document_type": "invoice",
  "confidence": 0.92,
  "raw_text": "...",
  "extracted_fields": {
    "vendor_name": "Office Depot",
    "invoice_number": "INV-2024-001",
    "invoice_date": "2026-01-15",
    "due_date": "2026-02-15",
    "total_amount": "250.00",
    "currency": "USD",
    "line_items": [
      {
        "description": "Printer Paper",
        "quantity": "10",
        "unit_price": "15.00",
        "total": "150.00"
      }
    ]
  },
  "field_confidence": {
    "vendor_name": 0.95,
    "invoice_number": 0.89,
    "total_amount": 0.98
  },
  "requires_manual_review": false
}

Step 5: Quality Check

Assess quality:

• •High Confidence (> 0.9) - Auto-process
• •Medium Confidence (0.7 - 0.9) - Flag for review
• •Low Confidence (< 0.7) - Require manual entry

Check for:

• •Missing required fields
• •Illegible text
• •Poor image quality
• •Incomplete document

Step 6: Post-Processing

• •Normalize Data - Standardize dates, amounts
• •Validate - Check logical consistency
• •Enhance - Add context from database (e.g., known vendor)

Error Handling

• •OCR Failed - Return error, suggest higher quality scan
• •Unsupported Format - Return error, list supported formats
• •Encrypted PDF - Request password or unlocked version
• •Large File - Split into pages, process individually
• •Poor Quality - Suggest rescan, adjust DPI

File Type Support

Integration Points

• •Google Document AI - Primary OCR engine
• •Tesseract - Fallback OCR
• •invoice-parser (AP worker) - For invoice-specific parsing
• •data-validator (Data worker) - For validation

Models

• •OCR: Google Document AI or Tesseract
• •Classification: Claude Sonnet 4 or Gemini Flash
• •Field Extraction: Claude Sonnet 4 (when OCR confidence low)

Security

• •Validate file type and size before processing
• •Scan for malware (if uploaded by user)
• •Never store raw documents longer than needed
• •Redact PII from logs
• •Encrypt documents at rest

Performance

• •Target: < 10s for invoice OCR
• •Batch Processing: Process up to 50 invoices concurrently
• •Caching: Cache OCR results for 24h (in case reprocessing needed)

Invoke this skill as the first step when processing any scanned or PDF document.