Extract Text (GOT OCR)

How it fits

Typical backends

Caveats

• Parameter interactiongotocr_mode switches between plain text and layout-preserving markdown/LaTeX output; choose the formatted mode for tables and formulas where structure matters.
• Parameter interactiongotocr_box restricts reading to a rectangular region and gotocr_color restricts it to a drawn box of a named colour; leaving both unset reads the whole frame.
• Latencygotocr_crop splits very wide or stitched two-page scans into patches and merges the results, improving accuracy at the cost of extra passes. Generation length is bounded internally, so very long pages may be truncated.
• Fallbackdevice requesting CUDA runs the model on the GPU; without a compatible accelerator it resolves to CPU and continues at CPU latency. The weights run in single precision either way, because GOT-OCR2's decoder overflows to NaN in half precision.
• LimitationThe output is a single String with no per-region geometry; use a detection-based OCR component when bounding boxes are required.
• State lifetimeEvery setting — the model, device, mode, crop, region, colour, and token cap — is read once at startup; changing any of them requires a redeploy.