Knowledge Base

The knowledge base is a structured FAQ system. You organize question-and-answer pairs into categories, and the AI can draw on this content to answer common questions accurately and consistently.

Why Use Knowledge?

Without a knowledge base, you'd need to either:

• Write a separate action for every question users might ask, or
• Put all the information into the stage prompt (which gets long and unwieldy).

The knowledge base gives you a clean, manageable way to add FAQ content. The AI automatically uses the right answers when it detects a relevant question.

How It Works

1. You create categories — groups of related Q&A pairs (e.g., "Billing FAQ", "Product Features"). Each category must have a Prompt Trigger that describes when it is relevant — this is what the classifier uses to match user questions to the category.
2. You add items to each category — specific question-and-answer pairs.
3. You enable knowledge on a stage.
4. During a conversation, the classifier detects when the user asks a question that matches a category's prompt trigger.
5. The relevant Q&A pairs are included in the AI's context, and it generates an answer based on them.

The key insight: Knowledge categories appear to the classifier as virtual actions — the prompt trigger describes the intent, just like an action's classification trigger. The classifier matches the user's question to the right category, and the AI then uses the specific Q&A items to form its response.

Creating Categories

Go to Design > Knowledge and click Create Category.

Category Fields

• Name — A descriptive label (e.g., "Return Policy", "Pricing FAQ").
• Prompt Trigger — A phrase that describes when this category is relevant. This is what the classifier uses to match user questions. For example: "The user is asking about returns, refunds, or exchanges." This field is required — without it, the classifier has no way to route questions to this category.
• Order — Controls the display order in the console.

Writing Good Prompt Triggers

The prompt trigger is how the classifier knows to route a question to this category. Be specific:

Trigger	Quality
"returns"	Too vague
"The user is asking about product returns, refunds, exchanges, or return shipping"	Good — covers the full range of related topics

Adding Items

Within a category, add question-and-answer pairs:

• Question — The question users might ask (e.g., "What is your return policy?").
• Answer — The accurate answer (e.g., "We accept returns within 30 days of purchase with the original receipt.").
• Order — Controls the display order.

Keep Answers Factual

Write answers that state facts clearly. The AI will use your answers as the source of truth and compose a natural response based on them — you don't need to worry about making answers conversational.

Using Knowledge on Stages

On the stage edit view:

Enable Use Knowledge — this makes all knowledge categories available for classifier matching in this stage.

Example

Category: Return Policy

• Prompt Trigger: "The user is asking about returns, refunds, or exchanges"

Items:

Question	Answer
What is your return policy?	We accept returns within 30 days of purchase with the original receipt. Items must be in original condition.
How do I get a refund?	To request a refund, visit your order page and click "Request Refund". Refunds are processed within 5-7 business days.
Can I exchange an item?	Yes, exchanges are available for items currently in stock. Visit our exchange portal or contact support.

When a user asks "How can I return something?", the classifier matches the Return Policy category, and the AI uses these Q&A items to form an accurate response.

Tips

• Use specific prompt triggers — Help the classifier accurately match user questions to the right category.
• Keep items focused — Each item should cover one specific question. Don't combine multiple topics in one item.
• Review periodically — FAQ content goes stale. Update answers when policies, prices, or processes change.
• Don't over-rely on knowledge — For complex scenarios that require multi-step interactions, use stages and actions instead.