Before you try it.

Only for installation and the model download. After setup, matching works with Wi-Fi off. The app and Ollama run on the same computer; order text stays there. Try the offline check.

It exports a CSV for manual import. It does not write directly into billing software. Check your software’s import format and the exported rows before billing.

Documented Hinglish shorthand, Hindi number words before units, and Devanagari text: पारले जी 5 वाला 2 पेटी is a synthetic example. The parser also normalises supported Devanagari, Gujarati and Bengali digits. Unfamiliar words may need correction; this is not general support for every language.

No. The catalogue, demo orders and benchmark are synthetic. No real shop has used the tool yet. The one exception is a small check on 11 order pieces typed by 3 friends (not shopkeepers): 0 wrong rows were auto-accepted, and it showed gaps the synthetic data missed (details). Prices and carton sizes are illustrative. Read the synthetic data card before interpreting the synthetic results.

Tested: MacBook Air M2, 8 GB RAM, Node.js 24, Ollama 0.40.3. The default model gemma4:e2b-it-qat is a 4.3 GB download (about 3.6 GB in memory); allow about 6 GB free disk and close other apps. Bigger models are too slow on 8 GB RAM. In the bake-off, a 6 GB model took up to 37 s per line. Windows and Linux are untested. Without Ollama, the app runs fuzzy matching only. Setup guide.

The code has an MIT license. Local inference has no per-order API bill. You provide the computer, storage and electricity; this site gives no measured operating-cost estimate. Read the license.

Yes. Download it in Ollama and start with --model <name>. Evaluate it on a locked test set before switching. Reported results apply to the named model and setup.

An amber row asks one short question, with a pre-filled suggestion when available. Confirm, change or skip it. Green requires all safety rules to pass. Code reads the numbers; the model only picks the product. See all 16 rules.