Extract Tables TATR

How it fits

Typical backends

Caveats

• Hard constraintThe bound `model` file slot holds a Table Transformer snapshot: either the whole-table detector or a structure-recognition variant that labels rows, columns, headers, and spanning cells. The two checkpoints emit different label sets, so downstream must resolve ids against the loaded model's label map.
• Parameter interactionThe detection checkpoint locates whole tables; bind the structure-recognition snapshot in the `model` slot when downstream needs row and column geometry rather than table outlines.
• I/O contractThe output is a downstream-resolved choice between [Rectangle<Double>] geometry-only boxes and [BoundingBox] boxes carrying class and confidence; the arm is fixed when the graph is deployed.
• CompatibilityOn the [BoundingBox] arm the class id is the bound model's own structure-element label, not a stable cross-model schema; pin the `model` slot snapshot and resolve ids via its label map.
• Parameter interactionscore_threshold at 0.0 keeps every raw detection including very low confidence ones; raise it to suppress spurious boxes.
• Fallbackdevice requesting CUDA resolves to CPU when no compatible accelerator is visible at startup; inference continues at CPU latency. An empty Image short-circuits to an empty list without invoking the model.
• State lifetimeAll settings are read once at startup; box coordinates are clamped to the image and zero-area boxes are dropped. Changing the model, device, or threshold requires a redeploy.