Lumen Acoustics Catalog Assistant

Source of truth: every fact comes from catalog.json , fetched fresh at runtime — nothing about the product line is hard-coded.
Grounding first: a deterministic retrieval layer finds the exact catalog field for your question before the model ever sees it; the model may only rephrase that fact, never add to it. If its rephrasing drifts, the app silently falls back to the verified catalog text.
On-device inference: when WebGPU is available, answers are generated locally by Qwen2.5-0.5B-Instruct via <a href="https://github.com/mlc-ai/web-llm" target="_blank" rel="noopener noreferrer">WebLLM — no prompt or catalog data ever leaves your browser. See sources.lock for exact pinned versions.
No coupons, ever: the catalog defines no discount codes or promotions beyond the listed sale price, so requests to invent one are refused.
Honest limits: questions outside the catalog get an explicit "I don't know", not a guess.

A small instruct LLM answers questions about the Aether product line fully on-device via WebGPU — every claim is grounded in catalog.json and cites its source.

Status starting… Backend detecting… Engine — TTFT — Throughput —

Loading the on-device model… this can take a little while the first time (weights are cached afterwards). You can already ask questions — they'll be answered instantly from the catalog while the model warms up.

Try: