Extract Text (Surya OCR)

How it fits

Typical backends

Caveats

• Hard constrainttask_name must be one of the supported task identifiers; any other value aborts startup. Use the boxes-bearing task so the recogniser returns geometry for the region arms.
• I/O contractThe output arm is fixed when the backend deploys: a text consumer pins the String transcript, a region consumer pins records with Polygon<Double> or Rectangle<Double> bounds. An empty Image short-circuits to the empty form of the resolved arm.
• FallbackThe recogniser runs at half precision on GPU and single precision on CPU. device requesting CUDA resolves to CPU when no accelerator is visible. A frame whose recognition raises is dropped to the empty form of the resolved arm so the stream keeps running.
• Costrecognition_batch_size, detector_batch_size, foundation_chunk_size, and foundation_max_tokens are VRAM knobs where zero means library default; foundation_model_quantize adds a quantisation pass.
• I/O contractThe detection and recognition checkpoints are mounted as single weights files through the component's file links (a detection weights slot and a recognition weights slot); the architecture and processor configuration travels with the component, so no checkpoint is fetched at runtime.
• Parameter interactiondisable_math left off can pick up stray inline math markup on pure-prose pages; separator joins lines on the String arm and is ignored by the region arms.
• State lifetimePredictors, device, and every other parameter are bound once at startup; change a parameter by redeploying.