diff --git a/src/components/labs/use-ocr-extract.ts b/src/components/labs/use-ocr-extract.ts
index 8ed9307d4..60cd70c9a 100644
--- a/src/components/labs/use-ocr-extract.ts
+++ b/src/components/labs/use-ocr-extract.ts
@@ -13,7 +13,8 @@
*/
import { useMutation, useQuery, useQueryClient } from "@tanstack/react-query";
-import { apiFetch, apiPost } from "@/lib/api/api-fetch";
+import { apiFetch, apiGet, apiPatch, apiPost } from "@/lib/api/api-fetch";
+import { ocrImageToText } from "@/lib/labs/local-ocr";
import { queryKeys } from "@/lib/query-keys";
import type {
OcrCapabilityDto,
@@ -49,7 +50,7 @@ export function useOcrCapability(enabled: boolean) {
});
}
-/** Upload + extract. Resolves with the proposed rows for the review screen. */
+/** Upload + extract (VISION mode). Resolves with the proposed review rows. */
export function useOcrExtract() {
return useMutation({
mutationFn: (file: File) => {
@@ -65,6 +66,55 @@ export function useOcrExtract() {
});
}
+/**
+ * TEXT mode (v1.18.10) — OCR the image IN THE BROWSER (tesseract.js), then POST
+ * only the extracted text. The raw image never leaves the device. Resolves with
+ * the same proposed-rows DTO the vision path returns, so the review/commit flow
+ * is shared verbatim.
+ */
+export function useOcrTextExtract() {
+ return useMutation({
+ mutationFn: async (file: File) => {
+ const text = await ocrImageToText(file);
+ return apiPost("/api/labs/ocr/extract", {
+ mode: "text",
+ text,
+ });
+ },
+ });
+}
+
+/** The local-OCR opt-in preference (read + toggle). */
+export interface LabsLocalOcrPref {
+ labsLocalOcrEnabled: boolean;
+}
+
+/** Read the current local-OCR opt-in flag. */
+export function useLabsLocalOcr(enabled = true) {
+ return useQuery({
+ queryKey: queryKeys.labsLocalOcr(),
+ queryFn: () => apiGet("/api/auth/me/labs-local-ocr"),
+ enabled,
+ staleTime: 60_000,
+ });
+}
+
+/** Toggle the local-OCR opt-in; invalidates the flag + the capability probe. */
+export function useUpdateLabsLocalOcr() {
+ const queryClient = useQueryClient();
+ return useMutation({
+ mutationFn: (labsLocalOcrEnabled: boolean) =>
+ apiPatch("/api/auth/me/labs-local-ocr", {
+ labsLocalOcrEnabled,
+ }),
+ onSuccess: (data) => {
+ queryClient.setQueryData(queryKeys.labsLocalOcr(), data);
+ // The toggle changes whether text-mode scanning is available.
+ queryClient.invalidateQueries({ queryKey: queryKeys.ocrCapability() });
+ },
+ });
+}
+
/** Commit the confirmed rows and invalidate the dependent reads. */
export function useOcrCommit() {
const queryClient = useQueryClient();
diff --git a/src/lib/labs/local-ocr.ts b/src/lib/labs/local-ocr.ts
new file mode 100644
index 000000000..03712de02
--- /dev/null
+++ b/src/lib/labs/local-ocr.ts
@@ -0,0 +1,62 @@
+/**
+ * v1.18.10 — in-browser (local) OCR for lab-report scans.
+ *
+ * Used by the TEXT-mode Lab-OCR path: a user whose AI provider cannot read
+ * images (ChatGPT-OAuth/Codex, a text-only model) OCR's the photo HERE, in the
+ * browser, and only the extracted TEXT is POSTed to `/api/labs/ocr/extract`.
+ * The raw image never leaves the device.
+ *
+ * tesseract.js (WASM) is LAZY-loaded via `await import(...)` so the multi-MB
+ * worker + WASM core + `deu+eng` traineddata land in the browser only when the
+ * user actually scans in text mode — never at app boot, and never for the
+ * higher-accuracy vision path. This mirrors the chart-defer discipline.
+ *
+ * Images only — tesseract.js cannot read PDFs, matching the vision path's
+ * image-only constraint for non-Anthropic providers.
+ */
+
+/** German + English: a German lab sheet needs `deu` for ä/ö/ü/ß; `eng` covers
+ * the many English analyte names and units that appear on the same sheet. */
+const OCR_LANGS = "deu+eng";
+
+export class LocalOcrError extends Error {
+ constructor(message: string) {
+ super(message);
+ this.name = "LocalOcrError";
+ }
+}
+
+/**
+ * OCR an image file entirely in the browser and return the extracted plain
+ * text. Throws `LocalOcrError` when the engine fails to load or recognise.
+ *
+ * A fresh worker is created and terminated per call. The traineddata is
+ * browser-cached after the first run, so the cost is paid once.
+ */
+export async function ocrImageToText(file: File): Promise {
+ // Lazy import — keeps tesseract.js out of the main bundle.
+ let createWorker: typeof import("tesseract.js").createWorker;
+ try {
+ ({ createWorker } = await import("tesseract.js"));
+ } catch {
+ throw new LocalOcrError("Failed to load the local OCR engine");
+ }
+
+ const worker = await createWorker(OCR_LANGS);
+ try {
+ const {
+ data: { text },
+ } = await worker.recognize(file);
+ const trimmed = text.trim();
+ if (!trimmed) {
+ throw new LocalOcrError("No text could be read from the image");
+ }
+ return trimmed;
+ } catch (err) {
+ if (err instanceof LocalOcrError) throw err;
+ throw new LocalOcrError("Local OCR failed to read the image");
+ } finally {
+ // Free the WASM worker; the traineddata stays browser-cached.
+ await worker.terminate().catch(() => {});
+ }
+}