naina 0.2.1

Read text from images on-device. OCR from naina's C++ core over FFI, with no network calls and no cloud API.

0.2.1

• README hero renders on pub.dev. GitHub raw serves SVG as text/plain with nosniff, so browsers refuse to render it as an image; the hero is now a PNG.
• Renamed the package doc directory to doc/ per the pub layout convention, and added LICENSE to the package root, both required by pub.dev.
• Eight further scripts: arabic, cyrillic, el, eslav, korean, ta, te, th. Use NainaLanguage constants; NainaLanguage.all lists them.

Changelog

0.2.0

First release of the Flutter binding.

• Naina.open() / readRgbSync() / close() over FFI to naina's C ABI

• readRgbInIsolate() helper for keeping reads off the UI thread

• tiny (~11 MB) and small (~54 MB) tiers

• Android: NDK build of the C++ core, ONNX Runtime from the official AAR, arm64-v8a and x86_64

• iOS: podspec against onnxruntime-c

• NainaModels.stage() fetches weights on the device. The Android core has no libcurl, so Dart downloads; the C core still decides the paths and verifies the hashes.

• language: selects the recognition alphabet — 'devanagari' reads Hindi, Marathi, Nepali and Sanskrit. An unknown value throws.

• Requires Android API 28+ (std::aligned_alloc).

Verified on an arm64 Android emulator: weights staged, then a real A4 page read as 33 lines at 0.992 mean confidence, matching the native build. Three on-device integration tests pass, plus 11 host FFI tests including a struct-layout check against the C compiler.

iOS is not verified: the podspec is written but has never been built or run.

Known limits from the core: weak handwriting; Arabic, Tamil, Telugu, Thai, Korean and Cyrillic unsupported (and wrong-alphabet reads return plausible wrong text rather than an error); text only — no layout or markdown in this binding yet.