OCR engines

Wox uses the operating system OCR implementation by default. It does not download an OCR model until an OCR consumer selects and downloads PaddleOCR v6 Small.

Each OCR consumer keeps its own ocr_model setting. The setting stores system or paddle_ppocrv6_small; model files themselves are shared.

The local layout is:

text

~/.wox/runtime/onnxruntime/1.27.0/<goos>-<goarch>/
~/.wox/runtime/sherpa-onnx/v1.13.4-wox.1/<goos>-<goarch>/
~/.wox/models/ocr/paddle_ppocrv6_small/

The PaddleOCR model downloader verifies the size and SHA-256 of every upstream ONNX/config file before atomically replacing the installed model. The PP-OCRv6 recognition character dictionary is generated from the verified inference.yml file.

ONNX Runtime is shared with dictation. The current signed Wox native release delivers ONNX Runtime together with sherpa-onnx; the runtime manager separates the installed files by runtime ownership. Native libraries intentionally remain loaded after first use because both ONNX Runtime and sherpa-onnx keep process-wide state.