feat(ocr): Tesseract.js as default scanner, AI as opt-in per port

The mobile receipt scanner now runs Tesseract.js in-browser by default —
on-device, free, and image bytes never leave the device. AI providers
(OpenAI / Claude) become a per-port opt-in for higher accuracy on
hard-to-read receipts.

- Lazy-load Tesseract WASM in src/lib/ocr/tesseract-client.ts (5 MB
  bundle dynamic-imports on first scan, not in main chunk)
- Heuristic parser src/lib/ocr/parse-receipt-text.ts extracts vendor,
  date, amount, currency, and line items from raw OCR text
- New port-scoped aiEnabled flag on OcrConfig (defaults false). Resolved
  flag never inherits from the global row — each port admin opts in
  independently
- Scan endpoint short-circuits to manual-mode when aiEnabled=false so
  the AI provider is never invoked unless the admin has flipped the
  switch
- Scan UI runs Tesseract first, then asks the server whether AI is
  enabled — uses the AI result only when its confidence beats Tesseract;
  network failures degrade gracefully to the local parse
- Admin OCR-settings form gains the per-port aiEnabled checkbox

Tests: 756/756 vitest (was 747) — +7 parser unit tests, +2 aiEnabled
config tests.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
Matt Ciaccio
2026-04-28 19:46:29 +02:00
parent 46937bbcb9
commit 2cf1bd9754
11 changed files with 693 additions and 38 deletions

View File

@@ -14,6 +14,7 @@ const saveSchema = z.object({
apiKey: z.string().optional(),
clearApiKey: z.boolean().optional(),
useGlobal: z.boolean().optional(),
aiEnabled: z.boolean().optional(),
});
export const GET = withAuth(async (req, ctx) => {
@@ -51,6 +52,7 @@ export const PUT = withAuth(async (req, ctx) => {
apiKey: body.apiKey,
clearApiKey: body.clearApiKey,
useGlobal: body.useGlobal,
aiEnabled: body.aiEnabled,
},
ctx.userId,
);

View File

@@ -27,9 +27,16 @@ export const POST = withAuth(
const mimeType = file.type || 'image/jpeg';
const config = await getResolvedOcrConfig(ctx.portId);
// Tesseract.js (in-browser) is the default. The server only invokes
// an AI provider when (a) the port admin has flipped `aiEnabled` on
// and (b) a key resolves. Otherwise the client falls back to its
// local Tesseract result.
if (!config.aiEnabled) {
return NextResponse.json({
data: { parsed: EMPTY, source: 'manual', reason: 'ai-disabled' },
});
}
if (!config.apiKey) {
// Manual-entry path — no OCR configured. Frontend will show the
// verify form with empty fields so the user can fill it in.
return NextResponse.json({
data: { parsed: EMPTY, source: 'manual', reason: 'no-ocr-configured' },
});