Optical character recognition is a quality-floor problem : no single engine handles every document type well. Printed financial statements suit one engine, handwritten signatures another, fixed-format invoices a third. Enterprises that standardise on a single OCR engine accept the floor of that engine's worst case ; the platform's approach inverts the relationship by treating engine choice as an internal implementation detail of the service.
- Tesseract 5.8 LSTM as primary. The latest LSTM-based release of the canonical open-source OCR engine. Strong on printed text, strong language coverage, well-tuned defaults.
- Alternate engine as fallback. A second engine kicks in when the primary returns below a confidence threshold or fails the layout-detection step. The dispatch is automatic ; callers see one result, not an engine selector.
- Integration-tested output parity. The CI suite confirms both engines produce comparable output across a corpus of representative documents ; the fallback path is exercised on every release.
- Legacy migration. Enterprises running a legacy OCR pipeline (commercial or in-house) repoint to the microservice with the same input contract ; accuracy at the boundary is at least the floor of the previous engine, often better.