ocr-backend-design

What problem does it solve?

It helps you design a reliable, configurable OCR pipeline for scanned PDFs by letting users choose between multiple OCR backends and by providing diagnostics when OCR is skipped or silently fails.

Core Features & Use Cases

• Pluggable OCR backends: Supports "ClaudeVision", "TextRecognize" (offline fallback), and "Custom" so you can hook any OCR/Vision API without hard-coding a single fixed engine.
• Mode-aware OCR triggering: Enables "Auto" (OCR only when plaintext looks too short), "Force" (OCR for all pages), and one-shot forced OCR for single extraction operations.
• Two-stage page rasterization: Rasterizes PDF pages via PyMuPDF first, then falls back to native Wolfram imports for robustness across environments.
• ClaudeVision image splitting strategy: Applies an upper/lower split with overlap to reduce Claude image-size limitations while preserving continuity of text.
• Actionable diagnostics: Returns fields like OCRAttempted, OCRCalled, OCRUsed, OCRFailReasons, and verbose tracing to distinguish “hook did not run” from “hook ran but failed.”
• Safety-minded reliability patterns: Avoids pitfalls like Quiet @ Check false negatives by using type checks and explicit error-string handling.

Quick Start

Ask for a new OCR backend template by setting SourceVaultOCREnable with the desired backend mode and implementing a hook Function that returns a string OCR result for each rasterized PDF page.