ai-bookkeeper reads receipt photos on your own machine. Earlier today it gained a second OCR engine, RapidOCR, which reads far more of the text than tesseract does. That changed where the errors are. With tesseract, two thirds of missed totals were lost in the OCR: the right digits were nowhere in the text. With RapidOCR, 45 of the 68 misses on 600 CORD train photos had the right total somewhere in the reading. The parser was dropping them.
So this run read those 45 one at a time. Most turned out to be small, ordinary failures:
- A colon.
Change:5,000did not read as an amount, because the reader will not take a number glued to a word. Worse,Pay Cash 50.000 Change:5,000sits on one line, and the cash line's last amount is the change. Five receipts. - A lost letter. One restaurant prints
DUEwith the amount right after it. RapidOCR dropped the E and the space (DU1,241,790), and our OCR cleanup then turned the D into a zero. Five receipts, all the same restaurant. - Words that look like a total and aren't.
TOTAL OTY 5PCS(Total Qty),Discount total,Total Discont,Ch ange Due. Four receipts. - Words that are a total and don't look like one.
Tatai...,Tatal,Tota........ These now count, but only when no real total line was read. - The total with its word lost. On several slips the word "Total" never came through, and the line that did say
TOTALheld the cash or a stray number. The receipt still shows the real amount twice: once on a line of its own, and once more as subtotal plus tax, or as cash less change. When two of those agree, and no check backs the read total, we now take the amount they agree on.
| photos | before | after |
|---|---|---|
| CORD train shards 1-3, RapidOCR (where the rules came from) | 515 / 583 | 539 |
| CORD validation, RapidOCR | 83 / 98 | 85 |
| US receipts (SRD), RapidOCR | 180 / 196 | 182 |
| CORD test, held out, scored once: RapidOCR | 80 / 95 | 83 |
| same, RapidOCR plus tesseract's four readings | 80 | 83 |
| same, the page | 81 | 84 |
In every set, no total went from right to wrong. Two taxes did. Each came after its total became right: the tax reader then met text that is still damaged. Both are written up in the eval note.
What it means. Once the OCR engine reads the text, a receipt parser has to cope with the habits of real printers and real OCR: glued punctuation, one printer's abbreviation, misspelt labels. Nine small rules covered 24 receipts. Of the 44 misses left, half are lost in the OCR and half are one-offs, such as an amount split over two lines or two kinds of separator in one number. The next useful measurement is not another rule. It is the full five-reading vote on these same 600 photos, which needs about 100 minutes of tesseract.
Data: CORD-v2 by Clova AI (CC BY 4.0) and the ExpressExpense sample receipts. No photos are copied into the repository. Results are estimates from these datasets, not a guarantee for your receipts. Check totals before you file them.