The knowledge base is the backbone of your assistant’s configuration. It organizes topics and interactions that guide responses to user queries. Each topic typically represents a single FAQ or use case and contains three key parts:

  1. Topic name: A concise, descriptive title summarizing the topic.
  2. Content: The main information, including sample queries and model answers.
  3. Actions: Follow-up workflows triggered by the topic, like sending SMS messages or handing off to a live agent.

A well-structured knowledge base ensures your assistant retrieves accurate information, performs tasks reliably, and delivers a consistent user experience.

Key components of a topic

Topic name

  • Clarity and precision: Use clear, descriptive titles such as “Opening Hours” or “Refund Policy” that reflect the topic’s content.
  • Analytics impact: Topic names influence search accuracy, dashboard analytics, and reporting.
  • Best practice: Avoid generic or overly broad names that can lead to misretrieval.

Content section

  • Core information: Add concise, actionable content that directly addresses user questions. Write in the tone you expect the assistant to use.

  • Example queries: Include 2–3 sample questions to improve retrieval accuracy.

    • Example for “Refund Policy”:
      • “How can I get a refund?”
      • “What is your return policy?”
  • Organize large topics: Break down complex topics into logical, smaller sections to ensure clarity and precision.

Actions section

The Actions section allows the assistant to perform workflows beyond delivering answers.

  • Common actions:

    • Sending SMS messages.
    • Triggering a handoff to a live agent.
    • Invoking custom functions (e.g., fetching live data).
  • Structured instructions: Clearly define what should happen when a topic is triggered.

    • Example: "If the user asks for emergency assistance, trigger a handoff to a live agent in the support queue."

Learn more: How to configure Actions

Best practices for building a knowledge base

Write clear, actionable content

  • Use direct language that is easy for users to understand.
  • Keep content concise and avoid ambiguity.

Add examples to guide retrieval

  • Include 2–10 sample user queries per topic.
  • Use these examples to clarify the scope of the topic for the assistant.

Example for “Payment Issues”:

  • User query: “Why is my card not working?”
  • Model answer: “Please check if your card details are correct or try using another card.”

Simplify and structure topics

  • Group related questions under a single topic if they share similar contexts (e.g., “Store Information” can include Store Hours and Store Locations).
  • Avoid overloading topics with unrelated information.

Plan for risky or critical scenarios

  • Identify high-risk situations (e.g., refunds, emergencies, escalations) and create clear KB topics or rules to handle them.
  • Add specific instructions or actions to ensure these scenarios are managed effectively.

Example:

  • Topic: Refund Requests
  • Action: “Trigger handoff to the billing team for refund-related queries.”

Define and test actions

  • Use the Actions field to link workflows like SMS notifications, function calls, or handoffs.
  • Test actions thoroughly to ensure they behave as expected in real scenarios.

Learn more: Actions Overview

Notes on global rules

  • Consistency: Use global rules to set tone, style, and boundaries for assistant behavior across all topics.
  • Examples in rules: Include sample responses to guide the assistant’s tone and length.

Example:

  • Rule: “Respond in a professional tone and limit responses to two sentences.”

Iterate and improve

  • Continuously test and refine your knowledge base based on real conversations. Monitor performance to identify gaps or improve clarity.
  • Update example queries and responses as needed to ensure the assistant aligns with user expectations.

By following these best practices, you’ll create a knowledge base that improves retrieval accuracy, enhances user interactions, and ensures reliable assistant performance.