Architecture

Lumen is a single Swift app with three main layers: the browser engine, the on-device model, and the local knowledge store. Each layer is either Apple-native or a thin wrapper around something Apple ships.

Layers

On-device AI

The model is a 4-bit quantized build of Meta’s Llama 3.2 1B Instruct, run through MLX Swift on the Neural Engine and GPU. At this size, the model fits comfortably on modern iPhones and returns short, cited answers in a few hundred milliseconds.

Sentence embeddings come from Apple’s NLEmbedding. Semantic search combines cosine similarity over those embeddings with FTS5 keyword ranking, then feeds the top passages into the model as context.

Storage

All content lives in a single SQLite database inside the app’s sandbox. The schema separates page bodies, sentence embeddings, topic labels, and generated summaries. An FTS5 virtual table sits on top for full-text search.

-- Rough shape of the content table
CREATE TABLE pages (
  id       TEXT PRIMARY KEY,
  url      TEXT NOT NULL,
  title    TEXT,
  body     TEXT,
  topic    TEXT,
  read_at  INTEGER
);

CREATE VIRTUAL TABLE pages_fts
USING fts5(title, body, content='pages');

The database is never exported over the network. Backups happen only through the phone’s own encrypted iCloud backup if you have one configured, which is opaque to Lumen.

Browser engine

The web view is Apple’s WKWebView configured for minimal ambient leakage:

• HTTPS-only upgrades on every navigation.
• Third-party cookies blocked by default.
• Mixed-content loads blocked.
• Tracker and fingerprint classification through ThreatDetector.

Lumen does not override or repaint the page DOM. What you see is what the site shipped, minus the instrumentation.