How Receipt Parsing Works: Inside Yomio's OCR Pipeline

How Receipt Parsing Works: Inside Yomio's OCR Pipeline

Yomio's current receipt workflow sends receipt images to Azure Document Intelligence's prebuilt receipt model. The backend parser maps the model's recognized receipt fields into Yomio's expense format, checks for missing or inconsistent values, and creates the expense record. The extraction can be wrong or incomplete, so a receipt scan should be reviewed before you rely on the saved data.

The receipt workflow

1. The app receives a receipt image

The user adds a receipt image through the app. Yomio sends it to the receipt-processing service, which calls Azure Document Intelligence with the prebuilt-receipt model. Microsoft documents the model's supported fields, inputs, versions, and languages.

2. The model returns structured fields

The model response can include merchant information, transaction date, subtotal, tax, total, currency values, and line items. Which values are present depends on what the model can identify in the receipt. The receipt parser reads the returned fields and converts them to Yomio's internal expense structure.

3. Yomio normalizes and checks parsed values

Yomio's parser handles dates and currency codes, normalizes merchant names where possible, and maps recognized item descriptions and prices into line items. It also records validation warnings for conditions such as missing totals, missing items, future or unusually old dates, and totals that do not reconcile with available line-item amounts. These checks can identify records to inspect; they cannot guarantee that every value matches the source image.

4. The app saves the expense

The backend uses the parsed result to create an expense and its available line items. Item categories can be supplied by the categorization workflow; if categorization is unavailable, items can remain in the default category. Review the expense and correct details that do not match the receipt.

What the workflow does not guarantee

We do not publish a controlled, representative benchmark for Yomio's receipt accuracy, line-item accuracy, or scan latency. NIST's AI measurement guidance says accuracy measures should use realistic, representative test sets and document the test method. Results vary with image quality, receipt layout, text, language, and the fields that a model can identify. We also do not claim that Yomio runs a second OCR provider in parallel or uses a custom receipt-trained OCR engine in the current receipt path.

Do not treat a confidence score or a validation warning as proof that a receipt is correct. For accounting, reimbursement, or tax records, compare important values with the original receipt and follow your organization's review process.

Why a receipt may need correction

• A photo is blurred, cropped, shadowed, or difficult to read.
• The receipt has faded text, unusual abbreviations, multiple tax lines, discounts, or a layout the model interprets differently.
• A value is visible, but the model assigns it to the wrong field or line item.
• The model does not return one or more fields needed for the expense record.

Take a clear image that includes the full receipt, then inspect the merchant, date, amount, currency, and any line items you need. Keep the source image available while you verify the saved expense.