The first fragrance AI that admits what it doesn't know.

Fragrance is a trust-heavy category. A hallucinated note kills credibility faster than any bug ever will. So Sniffbot Oracle ships with four non-negotiable rules — baked in the system prompt, enforced in retrieval.

The four rules

1. Never name a fragrance not returned by the tool retrieval. Hard rule.
2. Cite source for every factual claim — "notes per master record, last refreshed YYYY-MM-DD".
3. "I don't know" is a valid answer. Trained explicitly.
4. Launch with a 500-fragrance verified canon. Expand monthly, never beyond verified.

Sources

Parfumo, Basenotes, and AU retailers (MECCA, Libertine Parfumerie, Adore Beauty, Chemist Warehouse) for metadata and is-it-actually-available-here signals. Refreshed weekly via pg_cron.

Privacy

Your shelf and wear log never leave your account. ReasoningBank memory is per-user and non-transferable — Sniffbot does not train cross-user models on your data.

Changelog