Invoice Processing Workflow

Flagship demo workflow for RelayPlane. Extracts, enriches, and summarizes invoice data from images or PDFs using vision models.

Features

• Extract: Structured data extraction from invoice images using GPT-4o vision
• Enrich: Business intelligence analysis (payment terms, risk flags, GL codes)
• Summarize: Concise human-readable summary with actionable insights

Quick Start

Installation

# Install dependencies (from repo root)
pnpm install
pnpm build

Usage

import { invoiceWorkflow } from './templates/invoice/workflow';

const result = await invoiceWorkflow.run({
  apiKeys: {
    openai: process.env.OPENAI_API_KEY,
  },
  input: {
    imageUrl: 'https://example.com/invoice.png',
  },
});

console.log('Invoice:', result.steps[0].output);
console.log('Enrichment:', result.steps[1].output);
console.log('Summary:', result.steps[2].output);

Run Example

export OPENAI_API_KEY=sk-...
tsx templates/invoice/example.ts

Workflow Steps

1. Extract (Vision Model)

Uses GPT-4o to extract structured invoice data from images.

Input formats supported:

• Image URL: { imageUrl: 'https://...' }
• Base64: { imageUrl: 'data:image/png;base64,...' }
• Array: { images: ['url1', 'url2'] }
• OpenAI format: { content: [{ type: 'image_url', image_url: { url: '...' } }] }

Output schema:

{
  invoiceNumber: string;
  invoiceDate?: string;
  dueDate?: string;
  vendor: {
    name: string;
    address?: string;
    taxId?: string;
  };
  lineItems: Array<{
    description: string;
    quantity?: number;
    unitPrice?: number;
    amount: number;
  }>;
  total: number;
  // ... more fields
}

2. Enrich (Business Intelligence)

Adds financial analysis and business context.

Analysis includes:

• Payment terms (days until due, overdue status)
• Risk flags (duplicates, unusual amounts, missing info)
• Suggested GL codes for each line item
• Vendor history and patterns

Output:

{
  paymentTerms: {
    daysUntilDue: number;
    paymentStatus: 'due' | 'upcoming' | 'overdue';
  };
  riskFlags: Array<{
    type: 'duplicate' | 'unusualAmount' | 'newVendor' | 'missingInfo';
    severity: 'low' | 'medium' | 'high';
    description: string;
  }>;
  suggestedGLCodes: Array<{
    code: string;
    category: string;
    confidence: 'high' | 'medium' | 'low';
  }>;
}

3. Summarize (Human-Readable)

Creates a concise 2-3 sentence summary with key points and recommended actions.

Output:

{
  summary: string;
  keyPoints: string[];
  actionRequired?: string;
}

Example Output

═══════════════════════════════════════════════════════
📊 EXTRACTION RESULTS
═══════════════════════════════════════════════════════

Invoice Number: INV-2025-001
Date: 2025-01-15
Due Date: 2025-02-14
Vendor: Acme Corporation
Total: USD 6,540.00

Line Items:
  1. Web Development Services - USD 6,000.00
  2. Domain Registration - USD 540.00

═══════════════════════════════════════════════════════
💡 ENRICHMENT ANALYSIS
═══════════════════════════════════════════════════════

Payment Status: upcoming
Days Until Due: 29

⚠️  Risk Flags:
  1. [LOW] New vendor - first invoice from Acme Corporation

📝 Suggested GL Codes:
  Item 1: 5200 - Professional Services (high confidence)
  Item 2: 5100 - Technology Expenses (high confidence)

═══════════════════════════════════════════════════════
📝 SUMMARY
═══════════════════════════════════════════════════════

Invoice INV-2025-001 from Acme Corporation for $6,540.00 is due in 29 days for
web development services and domain registration. This is a new vendor with no
previous payment history. Recommended action: Verify vendor credentials before
processing payment.

Key Points:
  • First invoice from this vendor
  • Payment due February 14, 2025
  • Standard professional services categories

Architecture

Workflow Definition

The workflow uses RelayPlane's fluent builder API:

relay
  .workflow('invoice-processor')
  .step({ name: 'extract', provider: 'openai', model: 'gpt-4o', schema: InvoiceSchema })
  .step({ name: 'enrich', provider: 'openai', model: 'gpt-4o', dependsOn: ['extract'] })
  .step({ name: 'summarize', provider: 'openai', model: 'gpt-4o', dependsOn: ['extract', 'enrich'] });

Dependencies

• Extract → Enrich (sequential)
• Extract, Enrich → Summarize (join pattern)

Error Handling

The workflow uses RelayPlane's built-in error handling:

• Early abort on step failure
• Detailed error logs for debugging
• Retry logic for transient errors (rate limits, timeouts)

Testing

Run the test suite:

cd templates/invoice
pnpm test

Production Considerations

Vision Model Requirements

• GPT-4o or equivalent vision-enabled model
• Supports image URLs (public or pre-signed)
• Supports base64-encoded images

Cost Estimation

Per invoice (assuming GPT-4o pricing):

• Extract: ~1,000 input tokens + ~500 output tokens
• Enrich: ~800 input tokens + ~400 output tokens
• Summarize: ~1,200 input tokens + ~200 output tokens
• Total: ~$0.05-0.10 per invoice (varies by invoice complexity)

Performance

• Average workflow duration: 10-15 seconds
• Parallel execution not currently supported (sequential only)
• Vision step is typically the slowest (~5-8s)

Security

• API keys stored in environment variables
• Invoice images should be on secure infrastructure
• Extracted data may contain sensitive information (PII, financial data)

Customization

Custom Providers

Replace OpenAI with Anthropic, xAI, or Local models:

.step({
  name: 'extract',
  provider: 'anthropic', // Change provider
  model: 'claude-3-5-sonnet-20241022',
  // ... rest of config
})

Custom Schema

Modify schema.ts to extract additional fields:

export const InvoiceSchema = {
  type: 'object',
  properties: {
    // ... existing fields
    customField: {
      type: 'string',
      description: 'Your custom field description',
    },
  },
};

Custom Enrichment

Modify the enrichment prompt in workflow.ts to add custom analysis logic.

Support

For issues or questions:

• GitHub: https://github.com/relayplane/relayplane-workflows
• Docs: https://docs.relayplane.com
• Discord: https://discord.gg/relayplane