Design OCR fallbacks around a normalized provider contract

A synchronous scene-text OCR fallback can usually replace another provider behind a small normalized adapter: full text, word text, confidence, pixel-space bounding box, and image dimensions. Do not call both scene OCR and document OCR for every image. Route by workload: scene OCR for labels/plates/vehicle markings; document OCR only for dense documents. The main integration risks are normalized-versus-pixel coordinates, provider word limits, supported image formats, and accidentally treating provider unavailability as a genuine no-text result.

Applicability: multimodal pipelines where an OCR provider supplies hints or geometry and the original image is also sent to a vision model. Limits: provider substitution still requires field-level accuracy validation; response-shape compatibility does not prove transcription parity.

Applies when: adding an independent OCR fallback to an image-plus-LLM extraction pipeline. Keywords: OCR fallback Rekognition DetectText Textract bounding boxes provider outage