Install

Python

pip install naina

Wheels are self-contained: ONNX Runtime is bundled, so there is nothing else to install and no system library to match.

Supported: CPython 3.9–3.13 on macOS (arm64, x64), Linux (x64, arm64) and Windows (x64).

import naina

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

for line in page.lines:
    print(f"{line.confidence:.2f}  {line.text}")

Node

npm install @jvoltci/naina

Inference runs off the event loop, so a read does not block the process.

import { read } from '@jvoltci/naina';

const page = await read('scan.png');
console.log(page.markdown);

Browser

Nothing to install — use the tool.

To embed it in your own page:

npm install @jvoltci/naina-wasm onnxruntime-web

See Browser; there is one deployment detail you cannot skip.

From source

Needs CMake 3.24+, a C++20 compiler, and ONNX Runtime.

git clone https://github.com/jvoltci/naina
cd naina
cmake --preset macos-arm64        # or linux-x64, windows-x64
cmake --build build/macos-arm64
ctest --test-dir build/macos-arm64

Backends are opt-in

NAINA_WITH_ONNXRUNTIME defaults to OFF. A build without it produces a working library with no inference backend, and the test suite still reports green because every test that needs one skips.

The presets enable it. If you configure by hand, pass -DNAINA_WITH_ONNXRUNTIME=ON, and set NAINA_REQUIRE_BACKEND=1 when running tests to turn those skips into failures.

Model weights

Weights download on first use and are cached under ~/.cache/naina/models (override with NAINA_CACHE).

Every file is verified against the sha256 in models/registry.yaml. A mismatch is an error, not a warning — a corrupted cache entry will not be used.

To pre-fetch, or to run air-gapped:

# Fetch ahead of time
python -c "import naina; naina.fetch(tier='small')"

# Then refuse all network access at runtime
export NAINA_OFFLINE=1