dump anything. ask it back in your own words. ranked by meaning. nothing uploaded.
Everyone has the "I know I saved that somewhere" problem. You copy ~50 things a day and lose most of them — a wifi password, a docker run incantation, a flight number, that one address. Every notes app's fix is the same: organize it. Folders, tags, a system you abandon by week two.
cubby asks for nothing. Paste anything. One keypress. No folders, no tags, no title. Later, ask for it back in your own words — not the words you saved it with. An open-weight embedding model (EmbeddingGemma, 300M) runs entirely in your browser tab and ranks what you saved by meaning, offline.
-
Fully local. Your dumps never touch a server. The model is a file the app downloads once (~300 MB) and runs in the tab. No account, no API key, no telemetry. Saving, searching and switching themes made zero network requests in my check; the only traffic is the model download on first visit and a small model config request at page load, both to Hugging Face. Checkable in the network tab.
-
Semantic + lexical hybrid search.
0.7 · cosine + 0.3 · lexical, so a paraphrased query finds its target and an exact keyword never loses to a weak semantic guess. Measured recall@3 of 20/20 on a fixed test set vs. keyword's 14/20 — full honesty table here. -
Offline-first PWA. Service worker caches the app shell + model weights; search keeps working on a plane.
-
Share-to-cubby. On Android, the OS Share sheet writes straight into cubby — no app opening, no pasting. (iOS Safari's Web Share Target support is weak; see RISKS.md R8.)
-
Five themes × light/dark, all WCAG AA–verified (
test/contrast.mjs). Terminal / paper / mono / cobalt / ember. -
Keyboard-first.
Cmd/Ctrl+Kto focus, arrows to navigate,Enterto copy,Escto clear. No mouse needed for the full dump → ask → copy loop.
5 themes × light/dark. Each pair passes WCAG AA.
dump → ask → copy, in under 5 seconds. Also the 10-second intro video.
Live: https://cubby-6mn8.onrender.com
Or run locally:
git clone https://github.com/ibrahim-iqbal/cubby-hf26-wk3.git
cd cubby
npm ci
npm run devThen open http://localhost:5173. First load downloads the model (~300 MB, one time, cached by the service worker thereafter).
Build + preview the production bundle:
npm run build
npm run previewEverything inside the dashed box lives on your device, forever:
%%{init: {'theme':'base','themeVariables':{'primaryColor':'#141619','primaryTextColor':'#e8e9ea','primaryBorderColor':'#2d3136','lineColor':'#7ee0b8','fontFamily':'monospace'}}}%%
flowchart LR
subgraph browser["your browser tab — nothing leaves here"]
ui[React UI<br/>dump / ask]
model[EmbeddingGemma 300M<br/>transformers.js]
store[(IndexedDB<br/>items + vectors)]
sw[Service Worker<br/>offline cache]
ui <--> model
ui <--> store
sw -.caches.-> ui
sw -.caches.-> model
end
hf[Hugging Face hub<br/>one-time model download]
hf -.first load only.-> model
user((you)) --> ui
Full detail: ARCHITECTURE.md (model choice, Matryoshka 256-dim truncation, hybrid-ranking math, failure/fallback ladder), SPEC.md (features, flows, data shapes, acceptance criteria).
- UI: React 19 + Vite + TypeScript, static site, no backend
- Model runtime:
@huggingface/transformersv3 (transformers.js) runningonnx-community/embeddinggemma-300m-ONNXwith q8 quantization - Storage: IndexedDB for items + 256-dim Float32 vectors;
localStoragefor settings (read before first paint to avoid theme flash) - Offline:
vite-plugin-pwa(service worker precaches shell + runtime-caches model weights) - Deploy: Render static site —
render.yamlblueprint, no env vars needed
render.yaml is a Render static-site blueprint: npm ci && npm run build, publish dist, SPA-rewrite every route to index.html (so /share and client routes work). On Render pick New → Blueprint, point at your fork, done. No environment variables, no secrets, no cost beyond Render's free static tier.
Any static host works — Netlify, Vercel, Cloudflare Pages, GitHub Pages (with a rewrite rule) — the build output is a plain dist/ folder.
Issues and PRs welcome. The repo has a full planning trail if you want to read before contributing:
docs/PLAN.md— scope, non-goals, day-by-day plan, decisions logdocs/SPEC.md— features, user flows, data shapes, acceptance criteriadocs/ARCHITECTURE.md— model, embeddings, storage, search pipelinedocs/UI.md— minimal + terminal aesthetic, 5 themes × light/dark, motion rulesdocs/RISKS.md— risk register with mitigations and kill-switchestest/queries.md— D4 recall measurement + honesty table
Local setup: standard Node 22+. npm ci && npm run dev.
Verification scripts in test/:
node test/contrast.mjs # WCAG AA check across all 5 themes × 2 modesFor retrieval-quality checks, open the dev console in a running build and paste the small harness in src/testdata/queries.mjs (D4 honesty table).
Guidelines:
- Keep the "nothing uploaded" invariant honest. New features must not add outbound network calls at runtime (model fetch at first load is the only one, cached permanently thereafter).
- Changes to the ranking math must be measured against
test/queries.md. - All 5 themes × light/dark must pass WCAG AA (
test/contrast.mjs).
(Stated plainly, same as in the DEV post.)
- No image/OCR capture — text only, by design for v1.
- No sync across devices. JSON export/import is the only portability — there's no server to sync through.
- Hybrid ranking narrows but doesn't erase cases where the semantic and lexical signals disagree. Shown as a known behavior, not hidden.
- The honesty table is 20 items — it's the fixed set this project commits to re-running, not a claim the ceiling is infinite.
- iOS Share-to-cubby is unreliable (Web Share Target support is weak on Safari). Falls back to manual paste; stated outright rather than implied.
- EmbeddingGemma by Google — the model that makes this whole thing possible.
- transformers.js by Hugging Face — in-browser ONNX runtime.
- Render — static hosting, zero config.
- The Hacktoberfest Open-Source AI Challenge — the forcing function.
MIT.

