Every Arabic word grows from a three-letter root (جذر) poured into a pattern (وزن). Wazn shows a learner that structure instead of asking them to memorise past it.
React · Vite · FastAPI · K2 Think
Search عَامِل (ʿāmil, "worker") and Wazn takes it apart:
- Root — ع م ل, the idea of work
- Pattern — فَاعِل, the active participle: the one doing the root action
- So — one who works → worker
That same root in other patterns gives مَعْمَل (factory), مُسْتَعْمِل (consumer), اِسْتِعْمَال (use), عَمَل (job). Wazn draws them as one tree, because that is what they are: one family, one root.
Learn the root and the pattern, and you haven't learned one word — you've learned the shape of dozens.
The morphology is verifiable, not generated. The tree, the letter-by-letter breakdown, and the word family all come from a hand-curated knowledge base. GET /api/analyze runs no language model — every response carries "source": "deterministic" and a trace of how it got there. A learner is never shown a hallucinated root.
The model earns its place elsewhere. On a separate /api/insights endpoint, K2 Think writes the explanation, the practice questions, and an example sentence — then a guardrail agent checks that output back against the knowledge base before it's shown. Every K2 step falls back to a template if it fails.
Splitting the endpoints is the point: the trustworthy path stays fast, and the generative path can take its time.
Click any leaf and the companion panel follows it. Three tabs.
 |
Details. Root + pattern = word, spelled out, with root and pattern letters colour-coded — red for the root that carries the meaning, teal for the pattern that shapes it. Below that: why the word means what it means, an example sentence written for it, and the other words built on the same pattern. Recognise the pattern once and you can read all of them. |
 |
 |
| Quiz. Recall the family from the root alone, with the tree's labels hidden. | Insights. Which module found the word, which agent wrote the explanation, whether the grounding checks passed. |
Start the backend first, then launch the frontend in a separate terminal. Requires Python ≥ 3.10.
cd backend && pip install -r requirements.txt
uvicorn app.main:app --reload --port 8000cd frontend && npm install
npm run devOpen http://localhost:5173 and search عامل.
CORS errors in the browser mean your frontend origin is missing from the allowlist in backend/app/main.py.
search ──▶ GET /api/analyze ──▶ deterministic, no LLM
lookup · morphology · tree · quiz · guardrail
Insights ─▶ GET /api/insights ─▶ K2 Think, with fallbacks
explanation · quiz · sentence · guardrail · evaluation
backend/ is FastAPI: content is plain JSON in data/, loaded once at startup — no database. frontend/ is React + Vite: two views, no router, and a numbered CSS cascade where the order is load-bearing.