Skip to contents

Build an OCR-backed image scanner for the Data Shield input gate.

Source: R/input_gate.R
data_shield_ocr_scanner.Rd

Returns a function(content_image) -> list(action, ...) suitable for the image_scanner slot of .input_gate_scan() (or settings$data_shield_image_scanner). It OCRs the image with the optional tesseract package, then runs the extracted text through the shield's scan_prompt(). Images are otherwise a blind spot at edge 1 (the model sees pixels, not the redactable text inside them); this closes the gap for text baked into screenshots.

Usage

data_shield_ocr_scanner(shield, on_fail = c("block", "pass"), engine = NULL)

Arguments

shield

A DataShield engine (its scan_prompt() backs the scan).

on_fail

How an OCR hit is handled: "block" (default — immutable image cannot be redacted in place, so the whole turn is blocked) or "pass" (audit-only; log but let it through).

engine

Optional pre-built tesseract::tesseract() engine (language, whitelist, etc.). NULL uses tesseract's default English engine.

Value

A scanner function returning list(action = "block"/"pass", text=).

Details

This is opt-in (scheme A): the default data_shield_image_scanner stays NULL, so the host must wire this in explicitly. Reasons: OCR is a real per-image cost, it can false-positive, and tesseract is a Suggests dependency (a C library + language data, not installed by default). When tesseract is missing the scanner degrades to pass (never blocks on a missing optional dep) — so the image is treated as an accepted blind spot, matching the no-scanner default.