Agents

An agent is a specialized AI bot scoped to a set of document categories. Each agent has three configuration tabs: General Settings, Example Questions, and Extraction Rules.

Go to Agents → New Agent to create one, or click any existing agent to edit it.

---

Tab 1 — General Settings

Short Name

A technical identifier used internally. Rules:

• English letters only
• No spaces or special characters
• Must be unique across all agents in your organization

Example: supportbot, salesagent, onboarding

Human Name

The display name shown in the agents list and in escalation notifications.

Example: Support Agent, Sales Bot, Onboarding Assistant

Categories

Select which document categories this agent can search. You can assign multiple categories.

An agent only searches its assigned categories — it has no access to documents outside them.

Description

A plain-language description of what questions this agent handles. This text is passed directly to the LLM for routing decisions.

Be specific about the scope:

Good: Questions about product functionality, usage issues, access problems, error messages, and navigation in the user account.

Too broad: Customer support.

The more precise the description, the more accurately the orchestrator routes questions to this agent.

Fallback Behaviour

Defines what happens when the agent cannot find a relevant answer in its knowledge base:

• “I don’t know” — the agent responds that it couldn’t find the information. No escalation.
• Route to operator — the agent escalates the conversation to a human. The escalation includes all structured data extracted during the conversation (see Extraction Rules below).

---

Tab 2 — Example Questions & Explanations

Example questions teach the LLM which queries belong to this agent. They are used during routing when the orchestrator is deciding between agents.

Each example has two fields:

Example question — a realistic question a visitor might ask.

Explanation for LLM — a short explanation of why this question belongs to this agent.

Example

Question: “Do you issue certificates after the training?”

Explanation: “A question about certificates after training belongs to the TRAINING topic and should be handled by the training agent.”

Tips

• Add at least 3–5 examples per agent
• Include edge cases — questions that might seem ambiguous between two agents
• The explanation should reference the agent’s domain explicitly

You can add or remove examples at any time. Changes take effect immediately.

---

Tab 3 — Extraction Rules

Extraction fields define the structured data the agent collects from the conversation. When fallback is set to “Route to operator”, this data is included in the escalation payload as JSON.

Each field has:

Field	Description
Field name	The JSON key, e.g. email, order_id, incident_details
Description	What the LLM should ask for, e.g. “The order number from the confirmation email.”

If a visitor hasn’t provided a value for a required field, the agent will ask for it conversationally before proceeding.

Example fields

email         → The email address the customer used to register.
order_id      → The order number from the confirmation email.
incident_details → Technical details: error message, affected page, browser.

Escalation payload

When an agent escalates, your team receives the conversation along with extracted data in this format:

{
  "email": "[email protected]",
  "order_id": "ORD-4821",
  "incident_details": "404 error on checkout page, Chrome 121"
}

This means your team gets a structured pre-filled ticket instead of having to parse a raw chat transcript.