Scalable GenAI Invoice PDF Data Extractor

One-liner:
AI-powered backend system to extract structured data from multi-page invoices and PDFs and Images at scale, flag errors, and store results efficiently.

---

Table of Contents

• Features
• High-Level Architecture
• Dead Letter Queue (DLQ)
• Retry Strategy with Exponential Backoff
• Tech Stack
• Installation & Setup
• Scripts
• Environment Variables
• API Endpoints
• Sample Output
• Validation & Error Handling
• Scaling Workers & Concurrency
• Token Efficiency
• Postman Collection
• License

---

Features

• Batch & Async Processing: Upload multiple invoices; files queued in RabbitMQ.
• High Concurrency: Single worker can process ~100 files in parallel; scale horizontally with multiple workers.
• GenAI-Powered Extraction: Extracts vendor details, invoice metadata, line items, totals, and payment info.
• Error Detection: Flags mismatches in subtotal, sales tax, shipping, and total.
• Multi-format Support: PDFs, images, scanned, and multi-page invoices.
• Token-Efficient YAML Output: Saves 20–30% of AI token costs; converted to JSON for DB storage.

---

High-Level Architecture

Workflow Summary:

1. Users upload invoices via /api/file/upload.
2. Files queued in RabbitMQ.
3. Workers fetch id's and run GenAI extraction (parallel processing).
4. YAML output generated → converted to JSON → stored in MongoDB.
5. Validation checks applied; errors flagged.
6. Processed results accessible via API.

Dead Letter Queue (DLQ)

• Invoices or jobs that fail processing (e.g., unreadable files, persistent GenAI errors) are automatically routed to a Dead Letter Queue (DLQ).
• DLQ allows you to review failed jobs, retry them manually, or trigger alerts.
• Keeps the main RabbitMQ queue unblocked and ensures smooth processing of other invoices.
• Common use cases:
  • Corrupted PDF or image
  • Unsupported invoice format
  • Persistent AI extraction failure
  • or any other Error

Retry Strategy with Exponential Backoff

• Failed jobs are retried automatically using an exponential backoff strategy.
• Wait time increases after each failed attempt (e.g., 60s → 120s → 180s → 240s) to prevent overloading the system.
• Jobs exceeding the maximum retry count are sent to the Dead Letter Queue (DLQ) for manual review or alerts.
• Ensures smooth processing while handling temporary errors gracefully.

---

Tech Stack

Layer	Technology
Backend Framework	Express.js
Database	MongoDB
Queue / Async Jobs	RabbitMQ
Rate Limiting	Redis
Containerization	Docker & Docker Compose
AI / Data Extraction	Generative AI (Gemini)

---

Installation & Setup

🐳 Run with Docker

⚠️ Ensure .env includes your Gemini API key.

# Build and start containers
docker compose --build -d

# Access the app
http://localhost:3000

💻 Run Directly

# Start required services
docker compose up -d

# Start development server
npm run dev

# Start background worker
npm run worker

# OR start with PM2 for production
npm run start

---

Scripts

"scripts": {
    "dev": "node --env-file=.env --watch index.js",
    "worker": "node --env-file=.env worker/file.worker.js",
    "queue:flush": "node --env-file=.env scripts/amqp.flush.js",
    "start": "pm2 start ecosystem.config.cjs"
}

---

Environment Variables

GEMINI_API_KEY=<your-api-key>

AMQP_URL=amqp://user:password@localhost:5672
MONGODB_URI=mongodb://localhost:27017/invoices
REDIS_HOST=localhost
REDIS_PORT=6379

MAX_FILE_SIZE=200
PORT=3000
WORKER_CONCURRENCY=100

---

API Endpoints

1. Upload Files

• Endpoint: /api/file/upload
• Method: POST
• Request: Multipart/form-data (multiple files)
• Form Key: files
• Response:

{
 "message": "Files uploaded and queued",
 "file_ids": [array of file id's]
}

2. Fetch All Files (with filters)

• Endpoint: /api/file
• Method: GET
• Request Body Example:
• filters (status): pending, processing, processed, error
• flag: true or false (to fetch the flagged files having error in calculation or other)

{
  "flag": true,
  "status": "processed"
}

3. Fetch File by ID

• Endpoint: /api/file/{invoice_id}
• Method: GET
• FormData Example: Include file(s) to process if needed

---

Sample Output

items:
- name: "Activator CREAM 5-gal 5 gallons/pailt"
  quantity: 2
  rate: 269.00
- name: "AEB2461 PTFE 10\"x36YDS 5mil Glass Cloth Fabric / No Adh."
  quantity: 5
  rate: 64.00
- name: "Permabond 106 Cyanoacrylate 1-oz 10 bottles/case"
  quantity: 7
  rate: 139.00
- name: "Araldite 2014 HT Epoxy Paste GRAY 50ml 2:1 6 cartridges/box | 120 cartridges/case"
  quantity: 1
  rate: 719.00
payment_info:
  subtotal: 1187
  sales_tax_percentage: 8
  shipping_handling_cost: 50
  total: 1330.75
errors:
- "Subtotal mismatch: calculated 2550.00 vs declared 1187.00"
- "Sales Tax amount mismatch: calculated 204.00 vs declared 94.75 for sales tax rate 8%"
- "Total mismatch: calculated 2804.00 vs declared 1330.75"

Stored in MongoDB as JSON (YAML converted to JSON internally).

---

Validation & Error Handling

• Flags subtotal, sales tax, shipping, and total mismatches.
• Invalid invoices are stored with an errors array for review.

---

Scaling Workers & Concurrency

• Worker Concurrency: Default 100 files per worker.
• Multiple Workers: Run multiple workers to scale horizontally.
• File Size Limit: Default 200MB; configurable.
• Rate Limiting: ioredis ensures stable API usage.
• Tips: Monitor CPU/memory and RabbitMQ queues for optimal throughput.

---

Token Efficiency

• YAML formatting reduces AI token usage by ~20–30%.
• reduced characters by 30% (in current scenario) → faster output.
• Read more: How I Saved Millions in GenAI Token Costs

---

Postman Collection

• Import Postman collection from postman_collection.json.
• Base URL: http://localhost:3000
• Test endpoints: upload invoices, fetch all files, fetch file by ID.

---

License

This project is licensed under the MIT License.