Neural Notes — AI Search That Understands Meaning

Semantic note search running 100% in your browser — no server, no API key, your notes never leave the tab.

Day 001 of kb-daily-builds — one AI project a day.

What it does

Search "car trouble" and it finds your note about squeaking brake pads — zero shared words. Neural Notes embeds every note into a 384-dimensional vector using MiniLM running inside your browser tab via transformers.js, then ranks notes by cosine similarity to your query. A compare toggle shows plain keyword search side-by-side, so you can watch semantic search win. Notes persist in localStorage; nothing is ever sent anywhere.

_{Screenshot is auto-captured by CI on every push — if it's missing, the workflow is still running.}

Try it

Live demo → — first load fetches the ~23 MB quantized model once, then it's cached. Try feeling stressed, car trouble, or money planning against the seed notes.

git clone https://github.com/kbipul/neural-notes.git
cd neural-notes
npm install
npm test        # 15 unit tests on the search core
npm run dev     # http://localhost:5173

How it works

your note ──▶ MiniLM-L6-v2 (ONNX, WASM/WebGPU) ──▶ 384-dim vector
                                                        │
query ──▶ same model ──▶ query vector ──▶ cosine similarity ──▶ ranked results

Three decisions worth stealing:

1. The model is a lazy import. The app paints instantly; transformers.js and the ONNX model load in the background with progress surfaced in the UI. If the download fails, the app degrades to keyword search instead of dying.
2. The search core is pure TypeScript — no React, no model, no I/O — so it's fully unit-testable. The embedder hides behind a one-method interface; tests use vectors directly.
3. Vectors are never persisted. localStorage keeps only text (tiny, private); embeddings are recomputed on load. Model versions can change — stale vectors are a silent-corruption bug waiting to happen.

Build notes — what I learned

Quantized q8 MiniLM is a quarter of the full model's size and, for note-search, indistinguishable in quality. The interesting failure mode was scoring: normalized MiniLM cosine similarities cluster roughly between 0.1 and 0.8, so a raw score of 0.45 feels low while actually being a strong match. Mapping (cos+1)/2 into a visual bar keeps the UI honest without pretending to be a percentage-match oracle.

The other lesson is architectural: putting the entire pipeline client-side isn't just a privacy story, it deletes three operational problems — no inference server to scale, no key to rotate, no user data to secure. For a whole class of enterprise tools (this app is day one of a series exploring them), "ship the model to the browser" is now a serious design option.

Stack

Layer	Choice
UI	React 18 + TypeScript 5
Inference	transformers.js (Xenova/all-MiniLM-L6-v2, q8 ONNX)
Build / test	Vite 6, Vitest
Hosting	GitHub Pages (static — there is no backend)

---

_{Built by Kumar Bipul · IT Director → AI/ML · github.com/kbipul}