UPI Parser AI

A pure Dart package for sanitizing financial messages and parsing UPI transactions through Groq.

Installation

For this monorepo app:

dependencies:
  upi_parser_ai:
    path: ../../packages/upi_parser_ai

For another local project, adjust the path to this package.

Usage

import 'package:upi_parser_ai/upi_parser_ai.dart';

Future<void> main() async {
  final raw = 'Paid Rs.500 to Swiggy via GPay A/c 1234 user@okaxis';
  final sanitized = PiiSanitizer.sanitize(raw);

  final client = GroqClient(apiKey: 'your-groq-api-key');
  try {
    final transaction = await client.parseTransaction(sanitized);
    print(transaction.toJson());
  } on LlmParsingException catch (error) {
    print('Could not parse transaction: $error');
  }
}

PII Sanitization Logic

PiiSanitizer.sanitize redacts:

• UPI IDs such as user@okaxis, name@ybl, and uppercase variants.
• Explicit account/card numbers such as A/c 1234, Card No. 9876, Account No: 1234567890, and ending in 4567.
• Masked numbers such as xxx4567 or ****9876.

Cloud requests should only receive sanitized text.

Handling LLM Hallucinations

GroqClient.parseTransaction enforces JSON output and wraps failures in LlmParsingException. It rejects:

• Invalid JSON.
• Non-object JSON.
• Negative amounts.
• transactionType values outside expense, income, refund, transfer, or failed.
• confidence values outside 0.0 to 1.0.
• Empty provider, merchant, or category fields.

JSON Schema

{
  "amount": 450.0,
  "currency": "INR",
  "transactionType": "expense",
  "provider": "GPay",
  "merchant": "Swiggy",
  "category": "Food",
  "confidence": 0.92
}

Example

Run the package example from packages/upi_parser_ai:

dart run example/main.dart

Pass a real key to exercise Groq:

$env:GROQ_API_KEY="your_key"
dart run example/main.dart