A real-time AI learning companion for Google Docs that guides students through their mistakes — without giving away the answer.
Rethink AI is a Chrome extension that silently monitors what a student writes in Google Docs. When it detects an error — a grammar mistake, a flawed math equation, a factual inaccuracy — it nudges the student with a subtle location hint and opens a Socratic chat where an AI tutor guides them to discover the fix themselves.
The core principle: never give the answer, always guide the thinking.
Go to Release to start using Rethink AI :)
- Passive monitoring — watches for completed sentences or lines without interrupting the writing flow
- Error detection — sends document text to an LLM that identifies mistakes in writing, math, and science
- Location hints only — the extension only tells the student where to look, never what is wrong
- Socratic chat — a built-in chat panel in the popup lets students ask questions; the AI responds with probing questions, not direct corrections
- Subject-aware — the student selects Writing, Math, Science, or Other; prompts are tuned accordingly
- On/Off toggle — students can disable the companion at any time
- Works on all Google Docs — using the
/export?format=txtendpoint to read canvas-rendered text
Student types in Google Docs
│
▼ (debounce 1.8 s)
Content Script extracts document text
│
▼ chrome.runtime.sendMessage
background.js (service worker)
│
▼ POST /analyze
Backend (Node.js / Express)
│ LLM analyzes text → stores error internally
│ Returns only: { hasError, location }
▼
Extension tooltip: "⚠ Potential issue — look at: <location>"
│
▼ Student opens popup and asks a question
popup.js → background.js → POST /chat
│ Backend injects hidden error into system prompt
│ LLM replies with Socratic questions, never the answer
▼
Chat reply shown in popup
The backend detects the actual mistake and keeps it completely hidden from the extension. Only a vague location hint travels to the client. When the student opens the chat, the backend secretly injects the real error into the LLM's system prompt so it can guide the student without the extension (or the student) ever seeing the raw error description.
┌────────────────────────────────────────┐
│ Chrome Extension (Manifest V3) │
│ │
│ Content Scripts (7 files, ordered) │
│ ├─ config.js constants │
│ ├─ state.js shared state │
│ ├─ ui.js overlay + tooltip │
│ ├─ text-extraction.js /export fetch │
│ ├─ api.js storage + API │
│ ├─ analyzer.js debounce + loop │
│ └─ init.js listeners + MO │
│ │
│ background.js (service worker) │
│ └─ Proxies HTTP calls │
│ │
│ popup/ (Tailwind CSS) │
│ ├─ Toggle On/Off │
│ ├─ Subject selector │
│ ├─ Status banner │
│ └─ Socratic chat panel │
└───────────────┬────────────────────────┘
│ chrome.runtime.sendMessage
▼
┌───────────────────────────────────────┐
│ Backend (Node.js / Express) │
│ http://64.181.214.188:3000 │
│ │
│ GET /health │
│ POST /analyze → stores error │
│ POST /chat → injects error │
└───────────────┬───────────────────────┘
│ OCI SDK (instance principal)
▼
┌───────────────────────────────────────┐
│ OCI Generative AI │
│ Llama 3.3 70B Instruct │
│ us-chicago-1 │
└───────────────────────────────────────┘
| Layer | Technology |
|---|---|
| Extension | Chrome MV3, Vanilla JS, Tailwind CSS v3 |
| Service Worker | Chrome background.js (MV3 service worker) |
| Backend | Node.js, Express 4 |
| AI Model | OCI Generative AI — Llama 3.3 70B Instruct |
| Auth | OCI Instance Principal (no credentials in code) |
| Session store | In-memory Map (demo scope) |
| CSS pipeline | Tailwind CLI (npm run build:css) |
.
├── backend/
│ ├── server.js # Express API + OCI GenAI integration
│ └── package.json
├── extension/
│ ├── manifest.json # MV3 manifest
│ ├── background.js # Service worker — HTTP proxy
│ ├── content/
│ │ ├── config.js
│ │ ├── state.js
│ │ ├── ui.js
│ │ ├── text-extraction.js
│ │ ├── api.js
│ │ ├── analyzer.js
│ │ └── init.js
│ ├── popup/
│ │ ├── popup.html
│ │ ├── popup.js
│ │ ├── input.css # Tailwind source
│ │ └── tailwind.css # Generated — do not edit manually
│ ├── pictures/
│ │ ├── rethinkLogo.png
│ │ └── rethinkLogoBrain.png
│ ├── package.json
│ └── tailwind.config.js
├── api-contract.md # API source of truth
├── projectPlan.md # Full project plan + status
└── claudeMemory.md # Running architecture context
The backend runs on an OCI compute instance with instance principal auth. No credentials or
.envfiles are needed on that machine.
cd backend
npm install
node server.js # or: pm2 start server.js --name ai-companionThe server listens on port 3000. CORS is configured to accept chrome-extension://* and http://localhost:*.
-
Build the CSS (required after any popup change):
cd extension npm install npm run build:css -
Load in Chrome:
- Open
chrome://extensions - Enable Developer mode
- Click Load unpacked → select the
extension/folder
- Open
-
Navigate to any
https://docs.google.com/document/...URL and start typing.
Full specification: api-contract.md
| Endpoint | Method | Description |
|---|---|---|
/health |
GET | Liveness check |
/analyze |
POST | Detect errors in document text, return location hint |
/chat |
POST | Socratic response using stored session context |
POST /analyze — request
{
"sessionId": "uuid or null",
"subject": "writing | math | science | other",
"fullText": "...",
"newContent": "most recent sentence"
}POST /analyze — response
{
"sessionId": "uuid",
"hasError": true,
"location": "the third sentence"
}POST /chat — request
{
"sessionId": "uuid",
"message": "I don't see anything wrong with it"
}POST /chat — response
{
"reply": "What does the problem say about the direction of the force?"
}Google Docs renders text on a <canvas> element — there are no DOM text nodes to read. The only reliable approach is:
fetch(`https://docs.google.com/document/d/${docId}/export?format=txt`, {
credentials: "same-origin" // NOT "include" — causes CORS error on redirect
})Content scripts run in the HTTPS Google Docs context and cannot fetch HTTP endpoints directly. All backend calls are routed through background.js (the MV3 service worker), which is exempt from mixed-content restrictions.
Chrome's CSP blocks external CDN scripts on extension pages. Tailwind must be built locally and the output committed. Run npm run build:css from extension/ after any popup change.
Built for the University of Michigan AI Business Hackathon — theme: AI for Education.
The design philosophy aligns with the hackathon's goals of enhancing learning while remaining fair and transparent:
- The AI never completes the student's work
- It surfaces mistakes only after the student has written a complete thought
- Guidance is always question-based, keeping the student in the problem-solving loop
- The student can turn it off entirely
Think it through. Rethink it better.