> ## Documentation Index
> Fetch the complete documentation index at: https://docs.instaforge.app/llms.txt
> Use this file to discover all available pages before exploring further.

# AI brain

> Choose the AI that writes the answers, its personality, and how closely it sticks to what you taught it.

The **AI brain** page decides which AI service writes the assistant's answers and what it is told before every answer.

<Frame caption="AI brain: the service (1), the connection test (2) and the personality (3)">
  <img src="https://mintcdn.com/instaforge/HVDnlmyBMfIyfa_y/images/ai-brain.png?fit=max&auto=format&n=HVDnlmyBMfIyfa_y&q=85&s=e2a150b4cf50f6b0315130602060943f" alt="AI brain: the service (1), the connection test (2) and the personality (3)" width="1950" height="589" data-path="images/ai-brain.png" />
</Frame>

## Connect an AI service

<Steps>
  <Step title="Choose the service and model (1)">
    Pick the provider and type the exact model name from the provider's documentation. A small, fast model is usually best for customer chat: answers are quick and cheap.
  </Step>

  <Step title="Paste the API key and save it">
    The key is stored encrypted and is never shown again. Use the **client's own key** when you want AI usage billed to them directly.
  </Step>

  <Step title="Click Test connection (2)">
    You should see "Connected. The model replied." If not, check the key and the model name.
  </Step>
</Steps>

<Warning>
  Some providers process messages outside the EU or use free-tier data for training. The page warns you when you pick one of them. For customers in the EU, choose a provider and plan that doesn't.
</Warning>

## Personality and tone (3)

The **system prompt** describes who the assistant is and how it talks. Keep it short and concrete:

```text theme={null}
You are the friendly Instagram assistant for {{business}}.
Answer clearly and warmly in 1 to 3 short sentences.
Use simple words and at most one emoji.
If someone wants to book, offer the free times.
```

Safety rules are always added automatically: the assistant talks to one person, never reveals other customers' data, and answers business questions only from the knowledge base.

## Settings worth knowing

| Setting | What it does | Recommended |
| - | - | - |
| Creativity | 0 gives the same answer to the same question every time | 0 for customer service |
| Only answer from the knowledge base | Sends the fallback message instead of guessing | On |
| Send FAQ answers word for word | A clear FAQ match is sent exactly, without the AI | On for prices and policies |
| Messages remembered | How much of the conversation the AI sees | 6 to 10 |
| Remember each person | Keeps a short private profile: name, preferences | On |
| Fallback message | Sent when the answer isn't known | Friendly, offers a person |
| Hand-off message | Sent when someone asks for a person; the assistant then stays quiet in that chat | Say when the team usually replies |

<Tip>
  Set a **monthly AI budget** on the assistant's settings. When it's reached, the assistant stops calling the AI and you get an alert, so a busy month never brings a surprise bill.
</Tip>


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.