Receipt OCR is easy until the total is wrong

I like receipt OCR because it exposes the gap between “the text looks readable” and “the output is safe to use.” A clean grocery receipt can look trivial to a human. The model still has to decide what is a date, what is a tax line, what is a subtotal, whether a comma is punctuation or a decimal separator, and which tiny number actually matters.

The errors that matter are not evenly distributed

If OCR turns “THANK YOU” into “THANK Y0U,” almost nobody cares. If it turns $18.90 into $16.90, that is a different class of failure. The same is true for dates, VAT numbers, invoice IDs, merchant names and currency symbols.

That is why I would not judge receipt OCR by a single accuracy percentage. I would ask a more annoying question: did it get the fields I am going to rely on right?

What usually causes trouble

Thermal paper fades. Phone photos curl the page. Restaurant receipts mix bold totals with faint tips. Some stores print everything in uppercase. Others use condensed fonts, tiny line spacing or multiple columns. Add reflections from a glossy counter and a crumpled corner, and a perfectly reasonable OCR engine can start guessing.

The fix is often boring: crop tighter, shoot flatter, avoid shadows, keep the original resolution, and choose the right language. “Enhance” filters can help, but aggressive sharpening can also turn faint characters into weird edges that confuse recognition.

I would review five things manually

For expense or bookkeeping work, I would not read every line. I would check the merchant, date, total, currency and whatever identifier ties the receipt to a payment or expense claim. If those five are right, the OCR has probably saved time. If one is wrong, the original image stays the source of truth.

This is also where local OCR is useful. Receipts can expose where you shop, when you were there, what you bought, the last digits of a card, loyalty numbers, addresses and tax IDs. None of that needs to be sent to a random conversion server just to save a few minutes of typing.

Local does not mean magic

Running OCR in the browser changes where the file is processed. It does not make a blurry receipt sharper or turn handwriting into perfect structured data. The model still has limits, and a browser still has memory limits on very large batches.

What local processing does give you is a cleaner trust boundary: the image can stay on the device while recognition runs. For a handful of receipts, that tradeoff often feels better than uploading a pile of financial scraps to a service you know nothing about.