dsh-math-input

LIMITATIONS

已知限制

- **Model download**: first recognition triggers a ~7.2 MB download (encoder 3.4 MB + decoder 4.0 MB + vocab 4 KB); subsequent loads use IndexedDB cache. - **WebGPU auto-fallback**: when `webgpu` is selected but the browser doesn't support it (requires Chrome 113+), the engine automatically falls back to `wasm`. - **SharedArrayBuffer**: ONNX Runtime Web uses multi-threaded WASM when `Cross-Origin-Opener-Policy: same-origin` and `Cross-Origin-Embedder-Policy: require-corp` headers are present. Without them, it falls back to single-threaded — recognition still works but is slower. - **Handwritten vs printed**: the CoMER model is trained on the CROHME handwritten math expression dataset and optimized for handwriting. Screenshot OCR of printed formulas may underperform. - **Browser compatibility**: requires a modern browser with WebAssembly SIMD support. Chrome / Edge recommended; Firefox mostly works; Safari has limited support. - **Zero-token guarantee**: the plugin never calls `ctx.llm` — all recognition runs locally in the browser, with zero API costs.