Extract Text (PaddleOCR)

How it fits

Typical backends

Caveats

• I/O contractThe output arm is fixed when the backend deploys: a text consumer pins the String transcript, a region consumer pins per-line records with Polygon<Double> or Rectangle<Double> bounds. The polygon is PaddleOCR's native quad and the rectangle its axis-aligned bound.
• Hard constraintSelecting GPU via use_gpu requires CUDA at startup; there is no automatic switch to CPU for this setting.
• Parameter interactionpaddle_det_db_thresh, paddle_det_db_box_thresh, and paddle_det_db_unclip_ratio tune the DB detector; lower thresholds keep more candidate boxes and a larger unclip ratio widens them.
• Parameter interactionEach preprocessing stage has its own enable toggle and model slot. use_doc_orientation_classify with doc_orientation_model corrects whole-page 0/90/180/270 rotation, use_doc_unwarping with doc_unwarping_model dewarps curled or photographed pages, and use_textline_orientation with textline_orientation_model corrects per-line 0/180 orientation. Each adds latency; disable a stage when the source is already flat.
• State lifetimetext_detection_model and text_recognition_model select the detection and recognition models, and the three preprocessing model slots select their stages; each selected model is resolved into the shared model cache under paddlex_revision and loaded from there at startup. Changing any selection rebinds the engine at the next deploy.
• State lifetimeDisabling a preprocessing stage drops both its model from the cache scope and its stage from the engine; an enabled stage resolves its selected model into the cache. Selecting a model outside its slot's recognized set fails at startup rather than silently loading the wrong stage.
• State lifetimepaddle_lang, every model selection, and every preprocessing toggle are bound when the engine is constructed at startup; changing them requires a redeploy.