Barcode scanning
- Scan barcode in the food sheet reads UPC-A/E and EAN-13/8 live from
the camera with ZXing WebAssembly (bundled, precached, CSP allows
wasm), with photo and typed-digit fallbacks
- Lookups check saved foods first (offline), then local copies of USDA
FoodData Central Branded Foods and Open Food Facts, then the live
Open Food Facts API; API answers and misses are cached
- Server imports refresh in a low-priority child process: USDA when a
new archive is published, Open Food Facts weekly by streaming its CSV
export; each import builds a new table and swaps it in atomically
- Every available nutrient is kept: foods and entries gain optional
barcode (GTIN-14) and extra nutrients, shown under More nutrients and
included in JSON exports
- Settings shows the size, date and any error for each product copy
Label scanning
- Scan label sends a downscaled photo to the server, read with
PaddleOCR PP-OCRv5 (server detection, English recognition) on ONNX
Runtime CPU, about 1.3 s median
- Parser handles US/Canadian, linear, Spanish and EU per-100 g panels,
tilted photos, OCR misspellings and %DV cross-checks
- Models are pinned and checksum-verified at image build time
Container
- Runtime moves to Debian slim for ONNX Runtime's glibc build; other
platforms' binaries are pruned
- Smoke tests now read a fixture label inside the container
- README and Unraid docs cover disk use, network access and refresh
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>