Reviews sorted by tone and topic

Reviews get read in batches once a week, and the complaint that needed an answer today is third from the bottom of that batch. Half of the unhappy ones open with “thank you”, so keywords cannot tell them apart.

What is inside

1. A product review — the entry.
2. Parse the review — the model returns a tone, a topic and one phrase that captures it.
3. Where does this go — a three-way branch.
4. Unhappy goes to the quality chat, good goes into the collection, a question goes to support.

Tone is judged on substance, not politeness: “thanks, delivered fast, though the box was crushed” is negative about packaging. Keyword sorting reads that as gratitude, and the complaint is lost among the good reviews.

What to set up for yourself

• The topics in the topic field description: product, delivery, packaging, support, price. Yours are certainly different, and they become the axis of your statistics.
• The three outcomes. They differ for a reason: good ones accumulate, bad ones need a human now, disguised questions are not reviews at all. One shared “handle the review” branch would blur all of it.
• The quote. One phrase from the review inside the alert removes the need to open the original to see what it is about.

What to connect

A collection under Tables, a Telegram bot for the quality chat, and the Webhook target if reviews arrive from a marketplace or your site automatically.