naina

naina — reads any document, one C++ core, every language

Read any document. One C++ core, everywhere.

Try it online · Documentation · Architecture · Roadmap · Model weights

Use it in your browser now → No install, no upload, no account. PDFs and images, in ten scripts.

Scripts

`language`	Reads
(default)	Latin, Chinese, Japanese
`arabic`	Arabic, Persian, Urdu
`cyrillic`	Russian, Bulgarian, Serbian, Mongolian
`devanagari`	Hindi, Marathi, Nepali, Sanskrit
`el`	Greek
`eslav`	Ukrainian, Belarusian, Russian
`korean`	Korean
`ta`	Tamil
`te`	Telugu
`th`	Thai

page = naina.read("invoice.png", language="devanagari")

A tier picks model size; a language picks the alphabet. Detection and layout are script-agnostic and shared, so a language costs one 8 MB model rather than three.

Choosing wrong is silent. Read a Hindi page with the default alphabet and it returns fluent-looking Latin at ~0.75 confidence, not an error — confidence measures certainty within the model's own alphabet and cannot express "wrong alphabet". An unrecognised language value does raise.

v0.2.0 is out on PyPI and crates.io. The web app and docs are live. npm and pub.dev are blocked on account setup rather than on code — see the table below.

Packages

Package	Registry	What it does	State
`naina`	PyPI	Python, self-contained wheels	✅ published
`naina`	crates.io	Rust over the C ABI	✅ published
`@jvoltci/naina`	npm	Node, inference off the event loop	built; publish blocked, see below
`@jvoltci/naina-wasm`	npm	Browser, 143 KB brotli	built; publish blocked
`naina`	pub.dev	Flutter, FFI	built; Android verified on a device, iOS unproven
—	—	MCP server for LLM tools	works, in-repo

The npm publish fails with E404 on PUT @jvoltci/naina, which means the @jvoltci scope is not resolvable for the authenticated account — it must exist as an npm org or match the account's username. (Unscoped naina on npm is already taken by someone else, so the scope is not optional.)

import naina
import numpy as np
from PIL import Image

img = np.asarray(Image.open("invoice.png").convert("RGB"))

# One-liner: image -> markdown
print(naina.read(img))

# Or keep an Engine around
engine = naina.Engine(tier=naina.Tier.SMALL)
page = engine.read(img)
for line in page.lines:
    print(f"{line.confidence:.3f}  {line.text}")

Why

OCR accuracy is a solved 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 unwinnable and pointless.

The gap is distribution. Every existing tool is locked into one lane:

Tool	Lane	Cannot do
PaddleOCR	Python, server	No Node/Rust/browser/C ABI. Training framework first, huge surface
RapidOCR	Multi-language, as separate ports	Behaviour drifts between the Python, C++, Java and .NET versions
oar-ocr	Rust only	No Python/Node bindings, no shipped WASM
retto	Rust only, det+rec only	No bindings, no layout
client-ocr	Browser only	No server, no native
ML Kit (Google Lens)	Mobile only, closed weights	Cannot self-host; 5 scripts only
MinerU / marker / docling	Python, GPU-leaning	Licence traps, heavy installs

Nobody ships one engine that runs identically everywhere. This is llama.cpp's playbook applied to OCR: llama.cpp won on portability and zero dependencies, not on inference math.

No OpenCV. No pyclipper. No PaddlePaddle. The convex hull, minimum-area rectangle, polygon offset and contour tracing are ~450 lines of tested C++, because a 300 MB dependency tree would defeat the point of an 11 MB tier.

What you get

Three device tiers. A size axis, not a licence axis — every model naina ships is Apache-2.0 and safe for commercial use.

Tier	det	rec	layout	Total	Target	Charset
`tiny`	1.8 MB	4.5 MB	4.9 MB	≈ 11 MB	Browser, phone, Pi Zero	6,904 (CJK + Latin)
`small`	9.9 MB	21.2 MB	23.5 MB	≈ 55 MB	Laptop, Pi 5, mobile app	18,708 (50 languages)
`medium`	62.0 MB	76.6 MB	130.5 MB	≈ 269 MB	Server, desktop	18,708 (50 languages)

PaddleOCR ships no ONNX build of the small layout models, so naina converts them itself — byte-deterministically, and verified per-column against the Paddle original. Without that, layout would exist only at the 269 MB tier and an 11 MB browser build could not describe document structure. See tools/paddle2onnx_layout.py.

Every binding over one C ABI, so behaviour cannot drift between languages:

Binding	Status	Install
C / C++	✅ works	`naina.h` — the contract every other binding targets
Python	✅ works, unpublished	build from source; `pip install naina` once released
Node / TypeScript	✅ works, unpublished	build from source; needs a local toolchain
Rust	❌ v0.5	—
WASM / browser	❌ v0.4	—

Weights are mirrored, not borrowed. naina fetches from its own release, not from upstream hosting, so an upstream re-tag or deletion cannot break installs. Every file is pinned by sha256, so a corrupted or substituted download fails closed rather than producing silently wrong output. Provenance for each artifact is recorded in NOTICE and as a source_url in the manifest.

Benchmarks

Real measurements, not vendor claims.

End-to-end, tiny tier, Apple M3 Pro — rendered text fixture, 480×140:

Line	Recognised	Recognition conf	Detection score
1	`HELLO WORLD`	0.967	0.899
2	`naina 2026`	1.000	0.925

Reproduce: ctest --preset macos-arm64 -R test_ocr_e2e --output-on-failure

On the accuracy numbers everyone quotes. Vendors self-report 96.33% on OmniDocBench v1.6 while independent evaluation of the same benchmark tops out around 90.1%. naina will publish per-device numbers with the harness in-repo and the command to reproduce them, or publish nothing. A full benchmark matrix lands with v1.0.

Status

v0.2 — text spotting works end to end. The honest state:

Component	Status
C ABI (`naina_read`, page accessors, stage-level access)	✅
Model registry — manifest-driven, sha256-verified, tier fallback	✅
Detection — PP-OCRv6 det, DBNet decode	✅
Recognition — PP-OCRv6 rec, CTC greedy decode	✅
Geometry — convex hull, min-area rect, polygon offset, no OpenCV	✅
Page storage — pointer-stable, markdown + JSON	✅
Python binding	✅
Node binding	✅
ONNX Runtime backend	✅
NCNN backend	⚠️ compiles, but `FindNCNN.cmake` does not locate a brew install
Recognition batching (one strip per call today)	⚠️ correct but unoptimised
Layout analysis → structured markdown	❌ v0.3
WASM + browser app	❌ v0.4
Rust binding	❌ v0.5
Cross-binding parity enforced in CI	❌ v1.0
MCP server (`mcp/`, 2 tools, verified over stdio)	✅

13 C++ tests, 6 Python tests, 6 Node tests. CI builds on Linux (gcc + clang) and macOS arm64.

Not supported, deliberately: handwriting (PP-OCRv6 is weak at it and claiming otherwise would be dishonest), autoregressive VLM parsing, training, and chart/formula semantic extraction.

Install

Not yet on PyPI or npm — see the note at the top. Build from source:

cmake --preset macos-arm64           # or linux-x86_64, linux-arm64, windows-x86_64
cmake --build --preset macos-arm64
ctest --preset macos-arm64

Requires CMake ≥ 3.24, a C++20 compiler, yaml-cpp, libcurl, and ONNX Runtime.

naina ships no image decoder on purpose — it takes raw pixels. Use Pillow, OpenCV, sharp, or anything else that hands you a buffer.

Environment

Variable	Effect
`NAINA_CACHE`	Where weights are cached. Default `~/.cache/naina/models`
`NAINA_OFFLINE=1`	Disable network; use only what is already cached
`NAINA_REGISTRY`	Path to `registry.yaml`. Both bindings set this automatically

MCP server

An agent can read documents through naina directly:

{
  "mcpServers": {
    "naina": { "command": "npx", "args": ["-y", "@jvoltci/naina-mcp"] }
  }
}

Two tools: read_document (markdown) and read_document_detailed (per-line text, confidence, quads). See mcp/README.md.

Reading a page carries no session state, so the server is written stateless — which is what MCP spec revision 2026-07-28 formalised. Note that the current SDK (1.30.0) only negotiates up to 2025-11-25; the newer revision is a dependency bump away, not a rewrite.

Documentation

Architecture — the C ABI, backends, model registry
Roadmap — what ships when
Design spec — why naina is shaped this way
Contributing

The name

naina (नैना) means eyes in Hindi. The library reads.

It began as a face-recognition runtime under the same name. That work is preserved on the face-stack branch, and the engine it produced — C ABI, backend abstraction, manifest-driven model loader — is what made this pivot cheap.

Contributing

PRs welcome. See CONTRIBUTING.md. Open a Discussion for anything beyond a small fix.

License

Apache-2.0. Redistributed model weights are also Apache-2.0 — see NOTICE.

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

naina-0.2.1.tar.gz (333.4 kB view details)

Uploaded Jul 29, 2026 Source

Built Distributions

If you're not sure about the file name format, learn more about wheel file names.

The dropdown lists show the available interpreters, ABIs, and platforms. Enable javascript to be able to filter the list of wheel files.

naina-0.2.1-cp313-cp313-manylinux_2_28_x86_64.whl (10.8 MB view details)

Uploaded Jul 29, 2026 CPython 3.13manylinux: glibc 2.28+ x86-64

naina-0.2.1-cp313-cp313-manylinux_2_28_aarch64.whl (9.8 MB view details)

Uploaded Jul 29, 2026 CPython 3.13manylinux: glibc 2.28+ ARM64

naina-0.2.1-cp313-cp313-macosx_13_0_x86_64.whl (15.2 MB view details)

Uploaded Jul 29, 2026 CPython 3.13macOS 13.0+ x86-64

naina-0.2.1-cp313-cp313-macosx_13_0_arm64.whl (15.2 MB view details)

Uploaded Jul 29, 2026 CPython 3.13macOS 13.0+ ARM64

naina-0.2.1-cp312-cp312-manylinux_2_28_x86_64.whl (10.8 MB view details)

Uploaded Jul 29, 2026 CPython 3.12manylinux: glibc 2.28+ x86-64

naina-0.2.1-cp312-cp312-manylinux_2_28_aarch64.whl (9.8 MB view details)

Uploaded Jul 29, 2026 CPython 3.12manylinux: glibc 2.28+ ARM64

naina-0.2.1-cp312-cp312-macosx_13_0_x86_64.whl (15.2 MB view details)

Uploaded Jul 29, 2026 CPython 3.12macOS 13.0+ x86-64

naina-0.2.1-cp312-cp312-macosx_13_0_arm64.whl (15.2 MB view details)

Uploaded Jul 29, 2026 CPython 3.12macOS 13.0+ ARM64

naina-0.2.1-cp311-cp311-manylinux_2_28_x86_64.whl (10.8 MB view details)

Uploaded Jul 29, 2026 CPython 3.11manylinux: glibc 2.28+ x86-64

naina-0.2.1-cp311-cp311-manylinux_2_28_aarch64.whl (9.8 MB view details)

Uploaded Jul 29, 2026 CPython 3.11manylinux: glibc 2.28+ ARM64

naina-0.2.1-cp311-cp311-macosx_13_0_x86_64.whl (15.2 MB view details)

Uploaded Jul 29, 2026 CPython 3.11macOS 13.0+ x86-64

naina-0.2.1-cp311-cp311-macosx_13_0_arm64.whl (15.2 MB view details)

Uploaded Jul 29, 2026 CPython 3.11macOS 13.0+ ARM64

naina-0.2.1-cp310-cp310-manylinux_2_28_x86_64.whl (10.8 MB view details)

Uploaded Jul 29, 2026 CPython 3.10manylinux: glibc 2.28+ x86-64

naina-0.2.1-cp310-cp310-manylinux_2_28_aarch64.whl (9.8 MB view details)

Uploaded Jul 29, 2026 CPython 3.10manylinux: glibc 2.28+ ARM64

naina-0.2.1-cp310-cp310-macosx_13_0_x86_64.whl (15.2 MB view details)

Uploaded Jul 29, 2026 CPython 3.10macOS 13.0+ x86-64

naina-0.2.1-cp310-cp310-macosx_13_0_arm64.whl (15.2 MB view details)

Uploaded Jul 29, 2026 CPython 3.10macOS 13.0+ ARM64

naina-0.2.1-cp39-cp39-manylinux_2_28_x86_64.whl (10.8 MB view details)

Uploaded Jul 29, 2026 CPython 3.9manylinux: glibc 2.28+ x86-64

naina-0.2.1-cp39-cp39-manylinux_2_28_aarch64.whl (9.8 MB view details)

Uploaded Jul 29, 2026 CPython 3.9manylinux: glibc 2.28+ ARM64

naina-0.2.1-cp39-cp39-macosx_13_0_x86_64.whl (15.2 MB view details)

Uploaded Jul 29, 2026 CPython 3.9macOS 13.0+ x86-64

naina-0.2.1-cp39-cp39-macosx_13_0_arm64.whl (15.2 MB view details)

Uploaded Jul 29, 2026 CPython 3.9macOS 13.0+ ARM64

File details

Details for the file naina-0.2.1.tar.gz.

File metadata

Download URL: naina-0.2.1.tar.gz
Upload date: Jul 29, 2026
Size: 333.4 kB
Tags: Source
Uploaded using Trusted Publishing? Yes
Uploaded via: twine/6.1.0 CPython/3.13.14

File hashes

Hashes for naina-0.2.1.tar.gz
Algorithm	Hash digest
SHA256	`ae4e29880ef90f5492a9457531f1aa7a112f2332f61ce64bf1e6a98db1766638`
MD5	`8c49b7807ee7c7bb81c4dcb18c36a1f2`
BLAKE2b-256	`b43c4f584619aab193f597fdcbb22bca06e617b46ac9ff091add55ecffe3f5b2`

Algorithm	Hash digest
SHA256	`596fb12ab8bc0d38d50c40d43c20a92e91021115ddc3741fb6a890c24b431beb`
MD5	`64169db30d58190092bd7ee7c69b149c`
BLAKE2b-256	`2ff20c14fb40eff2b0f8127e3ebc498b8f13c54ed0dd01f7cfe5bd0acb247cc4`

Algorithm	Hash digest
SHA256	`fa9f4f57c2f104098b6d33e52dbaffda4bf3836d516e45933c4606340e115a76`
MD5	`ffd7539b51027527cc7c597e06594176`
BLAKE2b-256	`d0d3fb0aa89a830118234ceef3a11c37cf9ceb0fc4499e37cd48cc1a3ef6b0da`

Algorithm	Hash digest
SHA256	`75f9e4ec309ed2366dcca1b92f5e4bd0fd963e182f782922338fbb30c1f17a62`
MD5	`617684c0aa673df937fad234fef98d62`
BLAKE2b-256	`6918bf74add930c138ce33c98a43e068e8ca33767e97612d18a120d58de71c6e`

Algorithm	Hash digest
SHA256	`3b54a5c782cdea6bc26fe6a17ef88167cc3dd4b46930d6bba5469d17233ae968`
MD5	`97d0b5422886dd100f39eb1498cedf8d`
BLAKE2b-256	`3e640737c721ded775ba36b3ab3661173d816314db118feef339be54a08851de`

Algorithm	Hash digest
SHA256	`9882b8cd3bbaa87c8ee3f10504e99e0d0102bf7b6151b4a457eb4b84cadc3739`
MD5	`2fe4f52963588b04deaabed454c9a05b`
BLAKE2b-256	`a8c12fc79095de6a8bb517d1dedd760b4f145b294d151c6db6bc9f8ec2e3cbb9`

Algorithm	Hash digest
SHA256	`7e7bfe4801bf05d8f0c1bf9ca07e4fb2085a9494c7d6150d9e25d345a1e051e2`
MD5	`9dfafe1e4da137f475d5e450026c714d`
BLAKE2b-256	`dca968b9900a486b990bcc20c73588018b28b619f284f583f93439a903b65e14`

Algorithm	Hash digest
SHA256	`036864dcd3fcbc68df1622d113735f5fe3c4e58acd4d289514fe6c28ad880c5e`
MD5	`07b4a30bc67f8564981970546d870940`
BLAKE2b-256	`db2b7f3fb87bd3276d0e5d41e232df371f52aa7f1e29ba41542d4d0e7925c8be`

Algorithm	Hash digest
SHA256	`7231f7c0e3050b956036ab07a320e3032dc2872cc5095a36eff1557ea58f001d`
MD5	`2771f35c452a8be68f0be18d2b2cb061`
BLAKE2b-256	`e91bf7ef0a0beed19f7895b306651c4160ce53e618f3de6b638023df216826f3`

Algorithm	Hash digest
SHA256	`65aa175a8177a77ce6273e7e3126aa87d952f08e1f97babc7f4289d0790e00d8`
MD5	`116a4e5f1fe49497b5f2c1f74d970069`
BLAKE2b-256	`ed06429d3f40088ed6f9bf03b60680ac313fc9f796b1322b5e774caa1fd098ab`

Algorithm	Hash digest
SHA256	`05ec6eb580bfc561e4d2237099d3956b86c6b5c13ce7de37db807e690a2915a8`
MD5	`8b1bc4d1948d6cf1a46dbbae17572449`
BLAKE2b-256	`47ecf6c4224c7a7f87f681ab758dfa761a8da804038bda8fe644ef18d9bd8f11`

Algorithm	Hash digest
SHA256	`ad65a6b1903b4260e4c0b3b639187b50ece8aade179dbdeab11efdaa92400c2f`
MD5	`c84a0fbbd9d3f2d3729c533fff75a9bf`
BLAKE2b-256	`46b4ce7bbe57fa5067d9c5c82ca7446bf78412c382e9f30a97a81f1761f20dd1`

Algorithm	Hash digest
SHA256	`7263293115d7c4777626292d3692e9281b0024f09b87767d5fc7ff18171fdb21`
MD5	`62fa9f4e2e7f7e58e8fdef50b4fb551a`
BLAKE2b-256	`36c3fb8266defdf6a8617b0f60e209fb7b13b763ff3592a244b61cb75ab40868`

Algorithm	Hash digest
SHA256	`4ea5410c94914d3dc938b1c57272851168d5e27e0b25a8b8c58884f8fe3283a2`
MD5	`ce7d1e8ace30645ebaccc565f19f3ba5`
BLAKE2b-256	`7b4e51d24b0565fe161c51783830b6f1c5f85b44fd73a1f4fb8002692c878869`

Algorithm	Hash digest
SHA256	`97f5cbd160ac0d5d6e2c8cc02a210b666470fd8c366f4910b126e62062ef21d3`
MD5	`016c6abac6442c1a77e31dd8fe264cc2`
BLAKE2b-256	`6857288c9f0b9aafa4a4dfee4eb0c9a5a7a38960e2d47a670321c75aed56e125`

Algorithm	Hash digest
SHA256	`4ad8a94d87153e8538ffd1f3f162c10dfd714b9038e59c8e691c6afd638aae1a`
MD5	`24afbc7708cc299e6468d9b46e6a7a75`
BLAKE2b-256	`dee351471327385d05565e935c135bfebcf4586b498c68b0d51d35e3f123a055`

Algorithm	Hash digest
SHA256	`f3e9fea64982efbab34f92a110960f06831bc2792586118eacc2e003978afa6f`
MD5	`5d9c5e532268f1758f095c3adddd397c`
BLAKE2b-256	`9fdc81294a08d7a66bfd2eb4f26ad60544033f174c8fedb56269d01fe10e4dc1`

Algorithm	Hash digest
SHA256	`54a208f167178721463738b0273cc7c59792a7733e7f046b62f2e96e0409c7d5`
MD5	`ff730f2b2c0c166f2185afbbaf641a91`
BLAKE2b-256	`60f5d3ba04ef1d984ae6f7255d84e682b54d8a47dde0d96961e023649f12b396`

Algorithm	Hash digest
SHA256	`720790ac84587990b3664db26b54cd9c2c8f38041fe98ab0a3de835f79ee4800`
MD5	`c2bf58d77d326065476bbeef3f3b0840`
BLAKE2b-256	`aa351e364014d5f34321d0f1c95eac79d65a3bd9b59e51560b6aeab21a5fc3fb`

Algorithm	Hash digest
SHA256	`2d2fffd1a269348052a918b2297eafe4b04a93fc48de31d5a8e2a920b851891c`
MD5	`42eef3dab1c494e362eae322a0264679`
BLAKE2b-256	`00942efdc3d9d50daa7284d8566c56a7313b9d770d4aaf06408b6112517221f5`

Algorithm	Hash digest
SHA256	`f864337c3c192377ca4d2dcdeae1b5360397ef18e22a1ff5eead2193476a09d9`
MD5	`75c4a70e3c46eeb9a18d69a73a458c8b`
BLAKE2b-256	`0bc1a90cf3483fbd98eea98bb1a06a8aa1988a71d17c2ea170a04eece25f3b50`

naina 0.2.1

Read any document. One C++ core, everywhere.

Scripts

Packages

Why

What you get

Benchmarks

Status

Install

Environment

MCP server

Documentation

The name

Contributing

License

Project links

Key dates

1 maintainer

Credits

GitHub Statistics

License

Requires

Tags

Classifiers

Download files

Source Distribution

Built Distributions

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance

File details