MandarinShift

LocalShift already converts most digital-born files without trouble, but the scanned PDF path still lets me down: OCR on printed letters and reports often mis-extracts text, especially when the pages are in German, Japanese, or any other non-Latin script. I need you to dive into the existing Python codebase, pinpoint the weak links, and turn the OCR pipeline into a rock-solid, multilingual workhorse while keeping the whole app 100 % offline.

The project looks like this. A functional, test-covered backend written in Python orchestrates RapidOCR, ONNX Runtime, and a handful of helper utilities. Conversion logic is cleanly separated from the forthcoming PySide6 UI, so you will be working inside a fairly modular environment. Your task is to raise the reliability bar: scanned PDFs must convert to text-searchable PDFs (or plain-text output) with high accuracy, consistent encoding, and predictable memory/CPU use on Windows.

Typical bottlenecks you’ll address include:
• RapidOCR model selection and language packs for broad script coverage
• ONNX Runtime session tuning to balance speed and RAM on mid-range PCs
• Robust page segmentation so straight-text pages aren’t mistaken for images
• Failsafe error handling that clearly surfaces any unreadable pages rather than crashing the batch run

I expect clean, documented commits that pass the current test suite plus any new tests you add for edge cases. Unit tests should run from the existing pytest setup without hitting the network. Performance gains will be measured against the benchmark set in /tests/perf.

Before proposing changes, clone the repo, run the suite, and note where OCR accuracy drops. When you reply, highlight similar Python, PDF, OCR, or Windows desktop work you have completed and outline how you would approach multilingual model integration while preserving the offline guarantee.