Behaviour, policies, and safety

Business description, tone, six policies, FAQ, safety limits, escalation, Dry Run and Live.

Open Dashboard → Behaviour. Changes autosave.

Business and tone#

Business description — what you sell and who you are. Composed into the system prompt.
Tone — Warm, Efficient, Formal, or Playful.
Custom instructions — free text appended to the prompt. Keep it short.

There is no vector database. Policies and FAQ are included in the prompt directly.

Policies#

Six textareas: returns, refunds, shipping, exchanges, warranty, contact hours.

Each field is labelled: the bot may only quote what you write here. If a field is empty, the get_store_policy tool returns not-found and the model must not invent a policy.

FAQ#

Repeatable question + answer rows. Unlimited. Add the questions your team already answers by hand.

Safety limits#

Limit	What it does
Confidence threshold	Blocks low-confidence replies (check 3)
Groundedness threshold	Blocks replies the verifier cannot ground (check 4)
Max refund £	Blocks refund offers above this amount (check 9)
Max discount %	Blocks larger discounts (check 10)
Banned phrases	Blocks those strings (check 11)
Allowed link domains	Blocks links outside this list (check 12)
Max reply words	Warns (does not hard-block) when exceeded (check 15)

Escalation#

Set an email (falls back to the account email). Checkboxes: angry customer, legal/chargeback mention, low confidence, no order found, explicit request for a human.

When the widget shows the capture card, the visitor’s email and note are emailed to you with the transcript.

Dry Run and Live#

Dry Run (default on): generation still runs; visitors see the dry-run notice; Activity stores the full draft.
Live: master switch for the site. Off pauses the widget.

Turn Dry Run off only after you have read real conversations in Activity.