naina¶

Read documents. One C++ core, everywhere.

naina means eyes in Hindi.

What it does¶

Give naina a page. Get back the text, where each line sits, how confident it is, and the document's structure as markdown.

import naina

page = naina.read("invoice.png")
print(page.markdown)

# ACME Corporation

Invoice 2026-0417

| Item | Qty | Price |
...

Why it exists¶

OCR accuracy is a commodity. PP-OCRv6's weights are Apache-2.0, so naina runs the same models PaddleOCR runs and gets the same accuracy. Competing there is pointless.

What no OCR library ships is one engine that runs the same way everywhere. PaddleOCR needs Python and a 300 MB dependency tree. Browser libraries reimplement the pipeline in TypeScript. Mobile gets a different SDK again. Each of those is a separate implementation that can — and does — disagree with the others.

naina is one C++ core behind a stable C ABI, with thin bindings on top:


Python	pybind11
Node	N-API
Browser	the same core compiled to WebAssembly
Rust	over the C ABI (planned)
Flutter	FFI to the C ABI (planned)

The bindings contain no algorithmic logic. Detection post-processing, quad rectification, CTC decoding, reading order and markdown assembly happen once, in the core.

What that buys you¶

Python, Node and Rust produce byte-identical output for a fixed backend, device and tier. Not "similar" — identical.

The browser is close but not bit-exact, and that is stated plainly rather than glossed: onnxruntime-web is a different build of ONNX Runtime with different kernels, so probability maps differ in the last few float bits. See what it cannot do.

No heavy dependencies¶

Convex hull, minimum-area rectangle, polygon offsetting and contour tracing are hand-written C++. There is no OpenCV, no pyclipper, no PaddlePaddle.

That is not purism. An 11 MB model tier behind a 300 MB dependency tree is not an 11 MB tier, and OpenCV compiled to WebAssembly would exceed the entire browser budget on its own. naina's WASM binary is 143 KB brotli.

Model weights¶

PP-OCRv6 (detection, recognition) and PP-DocLayout (structure), all Apache-2.0.

Every file is mirrored into naina's own GitHub release and pinned by sha256, with the upstream URL kept only as provenance. Upstream can rename, move or delete anything and naina keeps working.

Where to go next¶

Install — pip, npm, or nothing at all in a browser
Tiers — 11 MB, 54 MB or 269 MB, and how to pick
What it cannot do — read this before deploying anything