Receipt OCR Normalizer

OCR alone gives you a noisy text dump. Most receipts can be reconstructed with high confidence if you enforce the arithmetic constraint — line items must sum to subtotal, subtotal plus tax plus tip must equal grand total. Using that constraint as a validation step catches most OCR errors automatically.

1. Run the image through a vision-capable LLM with a JSON schema that includes every field. 2. Apply the arithmetic checks: sum(lines) == subtotal ± 0.02, subtotal + tax + tip == total ± 0.02. 3. If checks fail, re-prompt with the original image plus the failing calculation and ask for the single most likely OCR error. 4. On second failure, emit the row with a 'manual_review' flag and the reason. 5. Normalize vendor name against a known-vendors table to prevent 'MCDONALDS' / 'Mc Donalds' / 'McDonald's' duplicates.