Make recognized scan text searchable and save it in PDF copies

.github/workflows/ci.yml

@@ -27,6 +27,7 @@ jobs:
       # Everything, not --omit=dev: the libraries bundled into the app (pdf.js, KaTeX, the Markdown and OCR
       # code) and Electron itself are devDependencies here, so leaving them out audits almost nothing that ships.
       - run: npm audit --audit-level=high
+      - run: npm test
       - run: npm run test:chat-math
       - run: npm run test:thumbnails
       - run: node --test scripts/test-word-files.cjs scripts/test-network-security.cjs scripts/test-release-build.mjs scripts/test-claude-path.cjs
@@ -54,8 +55,9 @@ jobs:
       - name: Smoke test
         if: runner.os != 'Linux'
         run: npm run smoke
+      # Keep PR and temporary CI-run screenshots on the runner; only main retains its existing artifact.
       - uses: actions/upload-artifact@v7
-        if: always()
+        if: always() && github.event_name == 'push' && github.ref == 'refs/heads/main'
         with:
           name: smoke-${{ matrix.os }}
           path: release/smoke/

.gitignore

@@ -6,3 +6,4 @@ release
 release-test
 release-*/
 dev-app-update.yml
+public/ocr

README.md

@@ -183,7 +183,7 @@ Ask AI is optional and desktop-only. See [Using Ask AI](#using-ask-ai) for setup
 ### Pick up where you left off
 
 - **Recent library:** keep up to eight recently opened PDFs with thumbnail previews and the last page read; reopen
-  with one click, remove individual entries, or choose **Clear all**.
+  with one click, remove individual entries, or choose **Clear all**. Removing entries also removes their saved recognized text; **Clear all** removes all saved recognized text. You can also clear saved recognized text from Settings without removing documents.
 - **Startup restore:** reopen the last active document automatically when it is still in the recent library.
 - **Reading state:** remember each document's page, scroll position, zoom, rotation, sidebar visibility, and two-page
   mode.
@@ -249,8 +249,9 @@ On macOS, use `⌘` wherever this table says `Ctrl`. Press `?` with a document o
 
 ### Current limits
 
-- Automatic OCR adds a selection layer in the reader. The source PDF stays intact. Document-wide search,
-  signature-label detection, and whole-document AI questions use the PDF’s embedded text.
+- Automatic OCR reads scanned pages on your device using the scan languages selected in Settings. Recognized text becomes available for selection,
+  document search and Ask AI. Saving a copy adds a searchable text layer when the file can safely be rewritten.
+  Pages whose text the export font cannot represent stay available for recognition when reopened.
 - Ask AI uses extracted text, not page images, and includes up to about 400,000 characters of document text. Long
   documents may be truncated; AI answers can be wrong, so check the cited pages.
 - Folio opens PDFs up to 300 MiB and 10,000 pages. Webpage conversion uses a separate session, so pages behind a

THIRD_PARTY_NOTICES.md

@@ -17,6 +17,9 @@ they cover, and Electron's and Chromium's licenses are in the app's installation
 | pdf.js standard fonts: Liberation | Liberation font license (GPLv2 with font exception) | Red Hat, Inc. |
 | pdf.js CMaps | BSD-3-Clause | Adobe Systems Incorporated |
 | pdf.js ICC profile | CC0-1.0 | |
+| [Tesseract OCR](https://github.com/tesseract-ocr/tesseract) via [tesseract.js](https://github.com/naptha/tesseract.js) and `tesseract.js-core` | Apache-2.0 | Tesseract OCR authors, tesseract.js contributors |
+| Tesseract English language data (`@tesseract.js-data/eng`) | MIT (wrapper); model from `tessdata_best`, Apache-2.0 | tesseract.js contributors, Tesseract OCR authors |
+| [pdf-lib](https://github.com/Hopding/pdf-lib) and its dependencies (`pako`, `@pdf-lib/standard-fonts`, `@pdf-lib/upng`, `tslib`) | MIT, Zlib, 0BSD | pdf-lib contributors, Nodeca (pako), and others |
 | [React](https://react.dev) and React DOM | MIT | Meta Platforms, Inc. and affiliates |
 | [react-markdown](https://github.com/remarkjs/react-markdown), [remark-math / rehype-katex](https://github.com/remarkjs/remark-math) and their unified parsing dependencies | MIT and ISC | unified contributors |
 | [KaTeX](https://github.com/KaTeX/KaTeX), including its bundled fonts | MIT | Khan Academy and other contributors |

electron/ai-prompt.cjs

@@ -10,20 +10,24 @@ const instructions = [
   "Be concise. Use short paragraphs, lists, and light Markdown. Use LaTeX \\( ... \\) for inline math and \\[ ... \\] for display math, outside code blocks.",
 ].join("\n");
 
+function instructionsFor(document) {
+  return document.recognized === true ? instructions + "\nSome or all of the text was read from scanned pages by OCR and may contain recognition errors. Allow for that and flag uncertain names or figures." : instructions;
+}
+
 function chatRequest(model, document, messages) {
   const input = [
     { role: "user", content: JSON.stringify({ document }) },
     ...messages.map(({ role, content, context }) => ({
       role, content: context ? JSON.stringify({ question: content, selectedPassage: context }) : content,
     })),
   ];
-  return { model, instructions, input, store: false, stream: true };
+  return { model, instructions: instructionsFor(document), input, store: false, stream: true };
 }
 
 function claudePrompt(document, messages) {
   // The CLI reads one user prompt. JSON keeps document strings inside their own fields even
   // when the PDF contains literal closing tags or text that resembles a conversation.
-  return { system: instructions, input: JSON.stringify({ document, conversation: messages }) };
+  return { system: instructionsFor(document), input: JSON.stringify({ document, conversation: messages }) };
 }
 
 module.exports = { chatRequest, claudePrompt };

electron/ai.cjs

@@ -547,7 +547,7 @@ const MAX_CHARS = 1_500_000;
 function validRequest(req) {
   return req && typeof req.id === "string" && req.id.length < 100 && ["chatgpt", "claude", "openrouter", "custom"].includes(req.provider)
     && typeof req.model === "string" && req.model.length <= (req.provider === "openrouter" || req.provider === "custom" ? 256 : 99)
-    && typeof req.document?.name === "string" && typeof req.document.text === "string" && typeof req.document.truncated === "boolean"
+    && typeof req.document?.name === "string" && typeof req.document.text === "string" && typeof req.document.truncated === "boolean" && typeof req.document.recognized === "boolean"
     && Array.isArray(req.messages) && req.messages.length > 0 && req.messages.length < 200
     && req.messages.every((m) => m && (m.role === "user" || m.role === "assistant") && typeof m.content === "string"
       && (m.context === undefined || (m.role === "user" && typeof m.context?.text === "string" && Number.isSafeInteger(m.context.page) && m.context.page >= 1)))
@@ -609,10 +609,10 @@ function register(handle, send, { confirmLocal } = {}) {
     active.set(req.id, controller);
     const emit = (text) => send("folio:ai-event", { id: req.id, type: "delta", text });
     const ask = { chatgpt: askChatGPT, claude: askClaude, openrouter: openRouter.ask, custom: customEndpoint.ask }[req.provider];
-    const { name, text, truncated } = req.document;
+    const { name, text, truncated, recognized } = req.document;
     const messages = req.messages.map(({ role, content, context }) => context
       ? { role, content, context: { text: context.text, page: context.page } } : { role, content });
-    Promise.resolve(ask({ model: req.model, document: { name, text, truncated }, messages }, emit, controller.signal))
+    Promise.resolve(ask({ model: req.model, document: { name, text, truncated, recognized }, messages }, emit, controller.signal))
       .then(() => send("folio:ai-event", { id: req.id, type: "done" }))
       .catch((e) => send("folio:ai-event", controller.signal.aborted ? { id: req.id, type: "done" }
         : { id: req.id, type: "error", message: e.message || "Something went wrong.", code: e.code ?? "", link: e.link ?? "" }))

electron/custom-endpoint.cjs

@@ -143,7 +143,7 @@ function createCustomEndpoint({ fetch, resolve, confirmLocal, safeStorage, file,
       const passage = messages.at(-1)?.context;
       if (document.truncated && !passage) throw new Error("This document exceeds Folio’s whole-document limit. Select a passage and ask again.");
       const { instructions, input } = chatRequest(model, passage
-        ? { name: document.name, text: `[Page ${passage.page}]\n${passage.text}`, truncated: true } : document, messages);
+        ? { ...document, text: `[Page ${passage.page}]\n${passage.text}`, truncated: true } : document, messages);
       const body = JSON.stringify({ model, messages: [{ role: "system", content: instructions }, ...input], stream: true });
       const name = nameOf(endpoint.baseUrl), timer = AbortSignal.timeout(5 * 60_000);
       try {

electron/openrouter.cjs

@@ -71,7 +71,7 @@ function createOpenRouter({ net, safeStorage, file, writeAtomic }) {
       const passage = messages.at(-1)?.context;
       if (document.truncated && !passage) throw new Error("This document exceeds Folio’s whole-document limit. Select a passage and ask again.");
       const { instructions, input } = chatRequest(model, passage
-        ? { name: document.name, text: `[Page ${passage.page}]\n${passage.text}`, truncated: true } : document, messages);
+        ? { ...document, text: `[Page ${passage.page}]\n${passage.text}`, truncated: true } : document, messages);
       // Small-context endpoints can otherwise silently drop the middle of the document.
       const body = JSON.stringify({ model, messages: [{ role: "system", content: instructions }, ...input], stream: true,
         plugins: [{ id: "context-compression", enabled: false }] });