Launch

Every review classified and traced — and the adverse ones routed toward the dossier.

Tamats classifies imported customer feedback — adverse reaction, efficacy, texture, scent, packaging, shipping, price, or unknown when nothing in the lexicon matches — using a deterministic, bilingual keyword lexicon rather than a model's judgement call, and adverse-reaction terms always win the category and are always flagged high severity, on the principle that under-calling a reaction is the expensive mistake. A per-product, per-category rollup flags a spike when recent volume clears a floor and genuinely rises against the equivalent prior window, and a feedback item flagged as an adverse reaction can be routed straight into the same dossier section that assembles undesirable-effects data for an EU PIF. Ingestion is CSV import only today — the feedback schema already has room for a direct Shopify or Cafe24 pull, but nothing wires it up yet, so a store's own reviews still have to be exported and imported by hand.

Feedback that arrives after launch is the cheapest signal a brand has and the easiest to lose in a pile of star ratings — an adverse-reaction cluster buried in a spreadsheet is exactly the kind of thing that should have surfaced weeks earlier. Tamats reads every imported item through the same deterministic lexicon every time, so the classification doesn't drift between a good day and a bad one, and keeps a straight line from a flagged reaction to the regulatory record that eventually needs it.

What you do with it

Deterministic, bilingual classification

Each feedback item classifies by keyword lexicon, English and Korean, into one category with a severity — adverse-reaction terms always win regardless of what else is in the text, and an item matching nothing in the lexicon is honestly marked unknown rather than forced into a category.

Spike detection against a prior window

A per-product, per-category rollup compares a recent window's volume to the equivalent window before it and flags a genuine rise, with a floor on the recent count so a small jump doesn't read as a dramatic spike.

Routed into the dossier

A feedback item classified as an adverse reaction can be shaped directly into the undesirable-effects section of the EU PIF dossier — the same record, not a re-typed summary of it.

CSV import today

Feedback comes in through a CSV import wizard now; the underlying schema already distinguishes a CSV, Shopify, Cafe24 or manual source, but only the CSV path is wired up to actually create one.

What it computes over

• —The organisation's own imported feedback items and their text
• —A deterministic, bilingual keyword lexicon (English and Korean) with no model call
• —The organisation's own product records, for the per-product rollup
• —The timestamp on each imported feedback item, for spike detection

Where it stops

• Ingestion is CSV import only. The schema already has a place for a direct Shopify or Cafe24 pull, but nothing automatically ingests from those connectors yet, even though both exist as connected sources elsewhere in the product.
• Classification is a keyword lexicon, not sentiment analysis or a model judgement call — an item with no matching term is marked unknown rather than guessed at, and a sarcastic or ambiguous review can land in the wrong bucket the same way any keyword match can.
• Spike detection needs a minimum recent count before it reports anything at all — a small absolute jump in a low-volume category is deliberately not flagged as a spike.
• Routing a reaction into the dossier drafts a section for a human to review; it does not itself notify a safety assessor or trigger any regulatory filing.

Questions

Does Tamats automatically pull in reviews from my Shopify or Cafe24 store?

No, not yet — feedback ingestion is CSV import only today. The schema already has a source kind for Shopify and Cafe24, anticipating that pull, but nothing currently wires it up, so reviews from a connected store still need to be exported and imported by hand.

How does Tamats decide a piece of feedback is an adverse reaction?

Through a deterministic, bilingual keyword lexicon — terms like burning, rash, hives or their Korean equivalents always win the category over anything else in the text and are always marked high severity, because under-calling a real reaction is the expensive mistake to make here.

What happens when a review doesn't match any category?

It's classified as unknown rather than forced into the nearest bucket — the classifier never guesses a category it has no lexicon match for, which is the same honest-or-absent posture the rest of the product runs on.