AI Configuration

Fine-tune how the AI agent behaves, when it escalates to a human, and which features are enabled. All settings are in Project Settings in the admin dashboard.

AI Models

Choose the model that fits your needs. Available models depend on your plan.

Haiku

All plans

Fast and cost-effective. Great for most customer support use cases — quick answers grounded in your knowledge base.

Sonnet

Pro & above

Stronger reasoning for complex multi-step questions and technical support conversations.

Opus

Business

Most capable model for enterprise use cases requiring the highest accuracy on complex queries.

Temperature

Controls how creative vs. factual the AI's responses are.

  • Low (0.0 - 0.3): Factual, consistent answers. Recommended for customer support.
  • Medium (0.3 - 0.7): Balanced creativity and accuracy.
  • High (0.7 - 1.0): More creative and varied responses.

Default: 0.2. For customer support, we recommend keeping it low to ensure reliable, consistent answers.

Agent Tools (Agentic Mode)

When enabled, the AI can take actions during a conversation instead of just generating text. Toggle this in Project Settings → Support OS.

Knowledge Base Search

The AI searches your knowledge base on-demand for each question instead of loading everything upfront. More accurate for large knowledge bases.

Smart Escalation

The AI uses judgment to decide when to escalate instead of relying on keyword matching. Triggers notifications to Slack and email automatically.

Contact Collection

Saves visitor name and email to the conversation when voluntarily provided during chat, for follow-up communication.

Auto-Resolve

The AI automatically resolves conversations when the customer's question has been fully answered, keeping your inbox clean.

Escalation Rules

Configure when conversations are handed off to your team:

  • Escalation phrase:A phrase visitors can type to request human help (e.g., "speak to a human"). The AI also detects escalation intent from context.
  • Blocked topics: Topics the AI should refuse to discuss (e.g., competitor pricing, legal advice). The AI responds with your configured off-topic message instead.
  • Off-topic response: The message shown when a visitor asks about a blocked topic.

Feature Toggles

Enable or disable features per project in Project Settings → Support OS:

FeatureDescription
Source CitationsShow which knowledge base sources the AI used in its answer
Semantic SearchUse AI-powered search to find relevant knowledge base content (more accurate than keyword matching)
Agent ToolsEnable agentic AI with smart escalation and on-demand KB search
Conversation MemoryRemember returning visitors across conversations for personalized support
Learning LoopAI learns from resolved conversations to improve future answers

Canned Responses

Create pre-written responses that agents can quickly insert when replying from the dashboard. Manage them in Project Settings → Canned Responses.