Model & behavior settings
Two different screens decide how your assistant behaves, and it's worth knowing which does what:
- Settings → AI Keys decides which model answers. The model comes from your own API key — Promptly asks your provider what that key can run and you pick from the answer. See Choosing a model.
- Settings → General — this page — decides how that model behaves: how creative it is, how long its replies can be, how much of the conversation it remembers, and whether the widget is live at all.
Open the second one at Dashboard → Settings. The page is titled Settings, subtitled "Configure your AI model and chatbot behavior", and everything on it is saved with one Save Changes button at the bottom.
🖼️ [Image] — The Settings page showing the AI Configuration card with Temperature, Max Tokens and Max History.
Which model actually answers
Your credential carries a model per role, not one model for everything:
| Role | What it does | Where you set it |
|---|---|---|
| Chat | Writes the reply your visitor reads. | Settings → AI Keys → Edit models |
| Planner | Decides which lookups and tools to run first, and classifies the question. | same |
| Background | Asynchronous work — FAQ mining, knowledge enrichment. | same |
Chat and Planner must be models that support tool calling; Background doesn't have to be. The full story is in Model roles.
The legacy Model field
The AI Configuration card still shows a Model select with three OpenAI names (gpt-4o-mini, gpt-4o, gpt-4-turbo). It predates BYOK, it does not read the model list from your key, and you should leave it on the default gpt-4o-mini.
That default is treated as "no choice made", so your credential's own Chat model is used. Choosing gpt-4o or gpt-4-turbo instead sends that literal model name to whichever provider your key belongs to — harmless on an OpenAI key, an instant error on an Anthropic, Gemini, GLM, or Kimi key, none of which have a model by that name. If replies start failing right after you touched this field, set it back to gpt-4o-mini and save.
AI Configuration
| Field | Range / step | Default | What it does |
|---|---|---|---|
| Temperature | 0.0 – 2.0, step 0.1 | 0.7 | Creativity versus consistency. Low (0.0–0.4) gives precise, repeatable answers — the right choice for support. High (1.0+) is more varied and less predictable. |
| Max Tokens | 256 – 8192 | 1024 | Ceiling on the length of a single reply. |
| Max History (messages) | 1 – 50 | 10 | How many earlier turns of the current conversation are sent with each request. |
A few things these three don't do, which saves guessing later:
- Temperature only shapes the wording of the final answer. The internal planning step always runs deterministically, so turning temperature up makes the bot chattier, never more resourceful about finding data.
- Some models don't accept a custom temperature at all — OpenAI's reasoning models (
gpt-5, theo-series) among them. Promptly omits the setting for those rather than failing the call, so the slider quietly stops mattering. - Max Tokens is a ceiling, not a target. Long multi-product answers are additionally capped by the assistant's own composition budget, so pushing this much past ~2000 rarely changes an answer.
- Max History is pure cost. Every extra turn is re-sent, and paid for, on every message. Raise it only if your visitors genuinely refer back ("and the second one you showed me?"). It is not how the bot remembers previous visits — that's Cross-session memory.
Cross-Session Memory
The same page carries the Cross-Session Memory section, which lets the bot recognise returning visitors. It's off by default and has its own page: Cross-session memory.
Widget
The Widget section holds one control, and it's the biggest switch in the product:
- Widget enabled — on by default. Under it: "When disabled, your chat widget will not load on your website. The embed script will still work but render nothing."
Turn it off and a banner appears: "Your widget is currently hidden from all visitors." This is how you pause the bot everywhere without editing your site — the embed script stays where it is and simply renders nothing. See Embedding the widget.
🖼️ [Image] — The Widget section with the toggle off and the "hidden from all visitors" warning banner.
Saving
Save Changes persists the whole page. The System Prompt field is required — with it empty you'll get "System prompt is required" and nothing saves. On success you'll see "Settings saved successfully".
Only Owner and Admin roles can save here; agents can't. See Roles & permissions (RBAC).
Tuning tips
- Support or factual bots — Temperature
0.2–0.5, Max Tokens512–1024. You want the same correct answer every time, not a fresh one. - Sales or conversational bots — Temperature
0.6–0.8. Enough variation to sound human, not enough to improvise facts. - If answers feel padded, lower Max Tokens before you rewrite your prompt. Models fill the space they're given.
- If the bot loses the thread mid-conversation, raise Max History a little — but check first that the visitor isn't actually asking something your knowledge base never covered.
- If answers are wrong rather than badly worded, none of these settings will fix it. Go to How retrieval works and Adding knowledge instead.
🎬 [Video] — Lowering temperature and max tokens, saving, and asking the same question again to compare the two replies.