The bot reads the room.
Every turn.
Vectrant runs a lightweight sentiment classifier on every customer message — detecting frustration level and the underlying reason (stuck, repeating, denied, confused, impatient, surprised, upset). The chat prompt is then tone-shaped in real time: empathy when the customer is heated, brevity when they're impatient, decisiveness when they're confused.
Tone shaping, one turn at a time.
A short Haiku call (max 30 tokens) classifies each customer message into a frustration level and a one-word reason. The result feeds back into the same conversation's next prompt as a tone directive.
Levels and reasons
Levels: low, medium, high. Reasons: stuck, repeating, denied, confused, impatient, surprised, upset, none. The combination drives different prompt shapes — for example, "high + repeating" tells the bot "the customer asked this twice, lead with empathy and a fresh angle, do not re-read the same script."
Cheap by design
The classifier runs only when there is signal — short queries are skipped, fresh conversations are skipped (no history yet). Failures default to "low" so a bug never makes the bot weirdly emotional.
Visible in the dashboard
Frustration levels are logged per message and rolled up in the conversation review. Patterns become visible: "this product page generates 30% high-frustration first turns" is information you can act on.
Frustration Detection, running on your data.
30-minute walkthrough on your real catalog. No slideware, no demo data.