Extract Text (Florence 2)

How it fits

Typical backends

Caveats

• I/O contractThe output arm is fixed when the backend deploys: a text consumer pins the plain String transcript and runs the page-level OCR task; any region consumer runs the region OCR task and pins records with Polygon<Double> or Rectangle<Double> bounds.
• AccuracyFlorence OCR exposes no calibrated per-region score, so the confidence-bearing arms report a constant value; do not treat that field as a real confidence.
• Hard constraintThe selected checkpoint must be a Florence-2 model; a gated or private repo additionally requires an access token configured before launch.
• FallbackOn a CUDA accelerator with bfloat16 support the model runs in bfloat16, which avoids the NaN logits that float16 produces in this vision encoder-decoder's autoregressive decode while keeping the half-precision footprint; without bfloat16 or a compatible accelerator it falls back to float32 (CPU resolves there too and continues at CPU latency).
• LimitationThe polygon carries Florence's native quad and the rectangle its axis-aligned bound; both are in pixel space, and decoding is deterministic with no sampling.
• State lifetimeThe model and device are read once at startup; changing them requires a redeploy.