ai-bookkeeper reads receipt photos on your own machine and matches each one to a bank line. This morning its page learned to read Russian slips with a Cyrillic recogniser (post), but only if you picked "Cyrillic" from a menu. The default was Latin. On a Russian slip the Latin reader still gets the total, so nothing looks broken. It just loses the seller and the item names: it took the seller on 16 of 146 held-out slips, against 106 with Cyrillic picked. Most people would never think to change the menu.
So the page now picks the alphabet itself. The new default, From the photo, reads each photo with the Latin recogniser, decides from that text whether the slip is Cyrillic, and if so reads it again with the Cyrillic one.
The tell: lower case in the wrong place
When a Latin-only OCR model meets Cyrillic, it prints the nearest Latin shapes. Upper-case Cyrillic mostly has look-alikes among Latin capitals: А В Е К М Н О Р С Т Х. But seven Cyrillic letters look like Latin lower case: а е о р с у х. On a printed slip, where the Cyrillic text is set in mixed case or in fonts whose small letters stand out, these come out as lower-case Latin letters in the middle of capitals:
CyMMAfor СУММА (amount)nPHXOAfor ПРИХОД (sale)KACCOBbIAfor КАССОВЫЙ (cash register), with ы read asbI
English receipts glue words too, since OCR drops spaces, but the case runs the other way: ChickenFried, BevSubtotal, ShowusyourreviewonTripAdvisor. A lower-case letter followed by two capitals is rare. The rule counts those words. Three or more, and the slip is Cyrillic. Brand prefixes such as McDONALD'S, iPHONE and eBAY are set aside first, so a fast-food receipt with three McSomethings stays Latin.
How well it picks
The threshold was set on two development sets, then the rule was scored once on held-out sets:
| set | receipts | picked right |
|---|---|---|
| Russian slips, val (development) | 149 | 147 Cyrillic |
| US receipts, SRD (development) | 200 | 200 Latin |
| Russian slips, test (held out) | 151 | 148 Cyrillic |
| Indonesian receipts, CORD (held out) | 100 | 100 Latin |
| US, European and Asian receipts, WildReceipt (held out) | 472 | 470 Latin |
On the development slips the count ran from 3 to 30. No US receipt had more than 2. Every slip the rule missed is short, 5 to 43 words, so there are only a few words to count. Some of those readings do contain KACCA and HAC 5% (НДС, the VAT line), so a rule based on known words might catch them. We saw those words only by looking at the held-out misses, though, so building that rule now would mean fitting it to the test. It waits for a fresh set.
WildReceipt's two misses went the other way: Latin receipts taken for Cyrillic. Both had quantities glued to item names, 1xBTL and 10xTVA, where the x is the lower-case letter before two capitals. A quantity is not a word, so the rule now cuts a leading 1x before it looks. With that, WildReceipt reads 472 of 472, and no other set's count moves. Because the fix came from WildReceipt, that 472 is not a held-out result. This direction matters more than the other: a Latin receipt taken for Cyrillic is read twice, and the Cyrillic reading leads.
What a visitor gets
On the 151 held-out Russian slips, with the menu left alone:
| reading | totals right | seller taken | item names found |
|---|---|---|---|
| Latin (the old default) | 137 of 148 | 16 of 146 | 81 of 459 |
| From the photo (new default) | 136 of 148 | 106 of 146 | 379 of 459 |
That matches picking Cyrillic by hand. One total is lost, on a short slip the rule left on Latin. The rule only looks at text, so we also checked the browser's own OCR on 57 of the photos in headless Chromium. It picked right on all 57.
Cost: a Russian photo is read twice, about five seconds more in the browser. The Cyrillic model (about 8 MB, hash-checked) downloads only when a photo needs it. A batch of English receipts never fetches it.
A bug found on the way
The test for the command line's new hint ("this receipt looks Cyrillic; set RAPIDOCR_REC_MODEL") failed for a reason unrelated to the hint. receipts photo.jpg --rapidocr was not running RapidOCR at all. The CLI's flag reader removes each flag as it reads it, and the receipt command checked --rapidocr twice. The first check consumed the flag, and the photo went to tesseract. With --four-pass added it worked. The evaluations read photos through their own scripts or in that mode, so no measured number changes. Running --rapidocr on its own now does what it says.
Method, every number and how to reproduce: eval/receipt-script-2026-11-19.md.