AI & Content
AI Models
Choose text, voice STT/TTS, Live/Realtime, and document-reader models per role — with safe fallbacks.
Overview
Assistant → AI Models (/en/settings/ai-models) picks which models power each role in your store: customer chat, panel tools, content, document reading, embeddings, and voice.
All metered generation bills AI credits (see Billing). Platform Help chat is separate and does not use your store wallet.
Role groups (what to set)
- Customer / support chat — WhatsApp, web widget, and related support turns
- Panel / admin tools — in-app AI that edits store settings when you ask
- Content — blog, pages, and similar long-form generate
- Document reader — shipping rate sheets, Pages AI Fill, large extract jobs
- Embeddings — Knowledge Base indexing and search
- Voice — customer voice messages, the voice-call model and panel voice helpers (see below)
Voice section
All voice settings are in one Voice section on this page:
- Customer voice
- Voice messages: the Live / Realtime model that answers WhatsApp, Instagram, Messenger and chat widget voice notes. It listens, uses the same store tools as typed chat and speaks the answer. If it fails, the text AI answers and a TTS voice reads the reply. Below the models you choose the reply format (voice + text, voice only, or text only) and the spoken voice (gender, voice and a sample).
- Voice calls: model, voice and speaking speed for voice assistant calls from your store website, app and phone line. A box under the picker shows which live-call controls the chosen model supports in Inbox → Calls.
- Voice assistant — a link to Channels & integrations → Voice assistant & phone. The on / off switch, the Talk to us test call, language, opening hours, phone line, WhatsApp calls, recording and blocked callers are all on that page; call history is in Inbox → Calls. See Voice Assistant.
- Panel voice — speech-to-text → reply → text-to-speech for voice commands in your panel only. It does not change the customer voice.
Fallbacks and tips
- Each role can have a fallback model when the primary is rate-limited or unavailable
- Faster / cheaper models suit high-volume chat; stronger models suit complex support and document extraction
- Changing models does not rewrite your AI Instructions — update Assistant → AI Instructions separately
- Changing models also does not fix missing Knowledge Base facts or contradictory policies — see Improve store AI answers
- Voice cards show your transcribe preference; Live voice options apply only when they are enabled for your store
Panel Ask AI family picker
The header Ask AI composer and the dashboard chat (including Automations) can pin a reply family for that turn:
- Auto — picks among the published Default / Fast / Quality / alternatives for Dashboard & pages
- ChatGPT / Claude / Gemini / Grok — locks the conversation model to that family from the published catalog
- Perplexity — best for live web search. Results are interpreted by OrderInChat’s AI models (Auto or the family you selected), not by a Perplexity chat model.
Live web search is available only in your panel. WhatsApp and the storefront widget do not search the live web.
You can also type answer with Claude (or ChatGPT / Gemini / Grok) in the same message; that overrides the chip for that turn only.
