Extract Text (docTR OCR)

@mindee1 VersionExtract

Two-stage docTR OCR over each Image; emits a joined String transcript or per-line regions with Polygon<Double> or Rectangle<Double> bounds (optional confidence). Detection + recognition weights are mounted file links, loaded locally.

How it fits

Typical backends

OCR document pages from files and post-process the transcript with a text model.

Pull frames over HTTP, OCR them, and extract named entities from the recognised text.

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. Detection and recognition run identically regardless of arm.
Hard constraintdoctr_det_arch and doctr_reco_arch must be valid docTR architecture identifiers; an unknown value aborts at startup.
Hard constraintThe detection and recognition weights are mounted as file links and loaded into the architectures named by doctr_det_arch and doctr_reco_arch; each weights file MUST match its architecture or the state-dict load aborts at startup. Nothing is fetched over the network at runtime; a missing weights file fails fast.
LimitationdocTR is not rotation-robust by default; deskew rotated scans upstream before OCR. Per-line text joins words with single spaces and punctuation or CJK spacing is not renormalised.
Fallbackdevice requesting CUDA resolves to CPU when no compatible accelerator is visible at startup; inference continues at CPU latency.
State lifetimeAll settings are read once at startup; changing the heads, weights, or device requires a redeploy.

Versions

4c925bealatestdefault
Automated release
6/29/2026