Python

pip install naina

Read a page

import naina

page = naina.read("invoice.png", tier="small")

print(page.markdown)          # structured document
print(page.text)              # every line, reading order, no markup

for line in page.lines:
    print(f"{line.confidence:.3f}  {line.text}")
    print(f"  quad: {line.quad}")     # 4 (x, y) corners, source pixels

Reuse the context

naina.read builds a context, loads models and tears it down. For more than one page, keep one around — model loading is the expensive part.

with naina.Reader(tier="small") as reader:
    for path in pathlib.Path("scans").glob("*.png"):
        page = reader.read(path)
        print(page.markdown)

Layout regions

for region in sorted(page.regions, key=lambda r: r.order):
    print(f"{region.order:2d}  {region.kind:8s}  {region.bbox}")

 0  title     (129.0, 154.0, 766.0, 92.0)
 1  text      (128.0, 293.0, 247.0, 20.0)
 2  title     (126.0, 375.0, 103.0, 22.0)

Kinds: title, text, list, table, figure, caption, formula, header, footer, pagenum, unknown.

header, footer and pagenum are page furniture and are omitted from page.markdown — they are not part of the document's reading flow.

Raw pixels

naina ships no image decoder. read() accepts a path and uses Pillow when it is available, but you can pass pixels directly:

import numpy as np

rgb = np.asarray(pil_image.convert("RGB"))   # H x W x 3, uint8
page = reader.read_rgb(rgb)

This is the path to use with OpenCV, a camera feed, or a PDF you have already rasterised — no re-encoding round trip.

Stages on their own

boxes = reader.detect("scan.png")        # quads, no recognition
regions = reader.layout("scan.png")      # regions, no text

Useful for cropping to a region before recognising it, or for checking whether a page has any text at all before paying for recognition.

Configuration

Environment variable	Effect
NAINA_CACHE	where weights are cached (default ~/.cache/naina/models)
NAINA_OFFLINE=1	refuse all network access; fail if a weight is missing
NAINA_REGISTRY	path to an alternative registry.yaml

Threads

A Reader is not thread-safe. Give each thread its own, or serialise access. Inference itself releases the GIL, so other Python threads keep running during a read.