ai-bookkeeper reads receipt photos on your own machine and matches each one to a bank line by merchant, date and total. Since yesterday its command-line reader has handled Russian slips with a Cyrillic recogniser, and the seller rules now pick the shop on 117 of 146 development slips (post). The web page did not have that recogniser. Its OCR still loaded only the default Latin and Chinese model, so in the browser a Russian slip got its total but almost never its seller.
What changed on the page
Next to "also read with RapidOCR" there is now a RapidOCR alphabet choice: Latin (the default) or Cyrillic, for Russian, Ukrainian, Belarusian and Bulgarian receipts. Cyrillic swaps in PaddleOCR's East Slavic recogniser (Apache-2.0, 7.9 MB). The detector stays the same.
That model is not on npm, where the page gets its other models, so it comes from Hugging Face. Hugging Face answers cross-origin requests for it, so the page can fetch it directly. Two precautions:
- The URL names a fixed revision (a commit hash), not
main, so a later upload cannot change what the page loads. - The page computes the SHA-256 of both files (the model and its character list) and refuses them unless they match the hashes recorded when the model was first measured. A swapped file gives an error saying so. It is never used.
Your photo is still never uploaded. The model comes to your browser, and the reading happens there, in a Web Worker so the tab keeps responding.
Does the browser read as well as the command line?
The browser runs a JavaScript port of the Python OCR pipeline, so the question is whether the port holds up with a different recogniser. We read the 149 development slips both ways: the CLI's Python engine, and the page's own loader in headless Chromium (worker, canvas and all).
| reading | totals right (of 145) | seller taken as the merchant (of 146) | item names in the text (of 482) |
|---|---|---|---|
| command line (Python) | 139 | 117 | 412 |
| page (browser) | 139 | 116 | 412 |
The rows of text matched exactly on 87 photos. 97.2% of the Python reading's lines appear unchanged in the browser's. Most of the gap is the browser's JPEG decoder producing slightly different pixels from Python's, as we found on US receipts. That cost one seller name. The Latin model takes the seller on about 20 of these slips, so for a Russian receipt the choice matters.
What is next
The page asks you to pick the alphabet. It could pick it from the photo instead: read once, and if the text is full of Latin look-alikes of Cyrillic words (NTOF for ИТОГ), read again with the other model. That needs measuring before it ships. Method, hashes and how to reproduce: eval/receipt-ocr-ru-page-2026-11-19.md.