Inspiration

Reading on the web is noisy and fragmented. We constantly bounced between tabs to summarize, explain, translate, and then hunt for credible sources. We wanted a privacy‑first, on‑device copilot that lives where you read and still scales when needed. Chrome’s built‑in AI Task APIs made that vision finally practical.

What it does

Summarize, Explain, Rewrite, Translate, Proofread, and Write inline on any page (PDFs too via Reader).
Prefers on‑device AI for speed and privacy; falls back to cloud transparently when needed.
Adds citations automatically:
- Preserves citations the model returns.
- If missing, finds 3–5 references via Google Programmable Search (CSE), deduped with short “reason” notes.
Research helpers:
- Concept drift analysis (key claim, agreement, drift) against your notes.
- Quiz from selection for quick comprehension checks.
- Library (Hub) to review saved Notes, curated Suggestions, and Quizzes.
Polished UX: selection toolbar, optional Floating Action Button (FAB), keyboard shortcuts, theming, onboarding tour, and a clear footer tip:
- Tip: On‑device when available; falls back to cloud. Used: On‑device/Cloud — “Why cloud?” explains the reason.

How we built it

Client (Chrome Extension)
- Content script: selection toolbar, panels, FAB, structured result parsing, device→cloud routing.
- Page bridge (injected): talks to Chrome AI Task APIs; prewarm + readiness probe; progress events.
- Background/service worker: hotkeys (chrome.commands), context menus, Reader launcher, messaging.
- Popup: login/signup, settings (mode, theme, persona, target language, cite sources), links to Reader/Library.
- State: chrome.storage.sync for persistent settings; postMessage bridge for on‑device events.
Backend (Spring Boot)
- /api/v1/ai → Gemini fallback; resilient with retry/backoff and structured result extraction.
- /api/v1/sources/find → Google CSE references with progressive queries and dedup.
- /api/notes, /api/v1/reading/suggest, /api/v1/quiz/generate-from-text for the Library.

Example backoff (code):

// Exponential backoff with jitter
function backoffDelay(attempt, base = 900, max = 4500, jitter = 300) {
  return Math.min(max, base * 2 ** attempt) + Math.random() * jitter;
}

Challenges we ran into

Google CSE 403 (“unregistered callers”): fixed by enabling the API, using a server‑side key, and avoiding referrer‑only restrictions.
Loader completion after auto‑finding sources: centralized render so loader.success fires after all async steps.
Function naming drift: showComparePannel vs showComparePanel caused ReferenceErrors; added a safe default + alias.
Settings persistence: “Cite sources” reset on persona change; introduced a citeSourcesManual override with storage listeners.
Structured outputs vary (fenced/quoted/escaped JSON): built a forgiving parser to recover bullets/citations reliably.
On‑device availability: first‑run downloads and restricted pages required a clear cloud fallback story and user messaging.

Accomplishments that we're proud of

Device‑first, privacy‑respecting AI with transparent cloud fallback and friendly progress.
Automatic references that feel native: subject‑aware search, deduping, and concise “reason” notes.
A cohesive UX: consistent “Summary/Explanation + References” across inline and Reader, themed UI, keyboard accessibility, quick onboarding.
A useful Library: saved notes, curated suggestions, and quizzes that extend reading beyond a single page.

What we learned

Chrome AI Task APIs are powerful and fast when prewarmed; pairing them with a robust fallback yields real‑world reliability.
Small UX details matter: status text, “Why cloud?” transparency, and consistent headings boost trust.
Sync semantics in extensions are tricky—debounce writes, listen for storage.onChanged, and never overwrite user overrides.
Dedup + query planning turn “web search” into believable citations; domain + normalized title carry most of the value.

We tuned backoff parameters (t_0, \alpha, t_{\max}) to balance responsiveness and load: $$ t_n = \min!\bigl(t_{\max},\, t_0 \cdot \alpha^n\bigr) + \varepsilon,\quad \varepsilon \sim \mathcal{U}(0,\text{jitter}) $$

What's next for PageGenie

Performance/reliability
- Lightweight server cache for CSE (5–10 min TTL).
- Local telemetry pane (op, path, latency, retries) for debugging and demos.
- First‑run model download indicator for on‑device tasks.
Features
- Multi‑language UI; export/import of Notes/Quizzes; deeper PDF support.
- Local RAG over saved notes (offline vector store).
- Optional TTS for summaries; OCR for images/PDF figures.
- Per‑site policies and custom personas.
Ecosystem
- Pluggable cloud providers; enterprise SSO/policy controls.
- Sharable Library items with privacy controls.