AI assistant

Persona

• Name — defaults to “Aria”. Set anything you like.
• Tone — friendly, professional, playful, concise.
• Style notes — free-form description (“Always use British English. Never use exclamation marks.”).

Model

By default we use OpenAI gpt-4o-mini for chat and text-embedding-3-small for retrieval. On Pro and Scale plans you can opt into larger models for harder questions.

Safety & guardrails

Found at Dashboard → AI Assistant → Safety & guardrails.

• Safety level — strict (recommended for storefronts), balanced, open.
• Allowed topics — comma-separated allow-list.
• Forbidden topics — block list (e.g. “competitor names, refunds older than 90 days”).
• Refusal message — what the AI says when asked something off-limits.
• Allow emojis, allow discount codes — fine controls.

Prompt-injection defense

We wrap every customer message in delimiters and instruct the model to never follow commands embedded in customer messages, retrieval data, or product descriptions. You can’t make the AI leak business data, expose system prompts, or change role from a chat alone.

Tools

Toggle which tools the AI can call. searchKnowledgeBase and requestHandoff are always on. Connecting Shopify auto-enables product/order tools.