oq-ai-ocr 0.1.0

Document ingestion and OCR: byte-level format detection, text-layer-first reading, local PP-OCR engines, an escalation policy, and the eval harness that settles all of it with evidence.

oq-ai-ocr

Document ingestion and optical character recognition (OCR, reading text off a picture), by OrbitQube. It is everything around the recognition engine and not the engine itself: byte-level format detection, text-layer-before-OCR reading, the format readers, an escalation policy that decides when a model is worth calling, and an eval harness that settles every threshold with a measurement rather than an argument.

This is the Python implementation. A TypeScript one answers the same contract, so a result crosses between them unchanged.

AGPL-3.0-or-later.

Install

The core installs with no recognition engine and no model client at all, which is what lets a consumer prove no model is even present. Extras pull the engine and the document readers:

pip install "oq-ai-ocr[rapidocr,documents]"

• rapidocr adds PP-OCR on ONNX Runtime, for reading scans.
• documents adds the PDF, Word, PowerPoint, Excel and image readers.
• server adds an optional HTTP surface.

Use

from oq_ai_ocr import extract

result = extract(data, filename="statement.pdf")
print(result.text)       # the whole document in reading order
print(result.engine)     # which engine read it, exactly
print(result.warnings)   # coded, one per thing that degraded the reading

What it does, and does not

• Reads a document's own text first, and recognises pictures only when there is no usable text.
• Runs locally. No document is sent anywhere. Recognition models download once and cache.
• Never raises for a document problem. An unreadable page, an unavailable engine, a corrupt zip and a timeout come back as a degraded result carrying warnings, so one bad page does not lose the rest. It raises only for a programming problem.
• Says what it did. Every result names the engine that read it, marks each page as text-layer or recognised, and carries a coded warning for everything that degraded the reading.
• No threshold is a guess. Every number that decides behaviour comes from a run of the eval harness and names the run it came from.

Reading order can be wrong

Recognised text is never perfect. Check a figure that matters before you trust it.