The knowledge base (KB) is the brain of your agent. Each topic defines a single intent, how users might phrase it, what the agent should say, and what it should do next.

Well-structured entries keep answers clear, safe, and meaningful.

​
How retrieval works

The agent does not see the whole knowledge base every time. Instead, it uses retrieval-augmented generation (RAG) to find the best match for the user’s message.

  1. The retriever compares the user’s message to the name, sample questions, and content.
  2. It returns the closest few topics.
  3. The model selects one and generates a reply (and may trigger an action).

​
Types of knowledge base entries

We will now go through the broad categories that most knowledge base entries tend to fall into:

​
Simple FAQ

Used when the agent just needs to answer a question with no follow-up or action required.

​
Single turn

One message in, one message out.

  • Add a short, helpful reply in content.
  • Leave actions empty.

Example – Pet policy (single-turn)

​
Multi turn

Ask a clarifying question before giving the answer.

  • Use branching inside content to structure the reply.
  • Leave actions empty.

Example – Pet policy (multi-turn)

​
Handoff

Used when the agent should offer to connect the user to a human agent.

​
Offer

Agent replies, then offers to transfer.

  • content: Include the answer + offer.
  • actions: Trigger transfer_call only if user accepts.

Example – Spa booking

​
Direct

Immediate transfer with no agent reply.

  • Leave content empty.
  • Always run transfer_call in actions.

Example – Billing handoff

​
Conditional

Transfer based on user clarification, for example, group size, request type.

  • content: Ask a disambiguating question.
  • actions: Map each answer to the correct destination.

Example – Room reservations

​
SMS

Used when you want to follow up with a link or info through text.

​
Offer

Offer to send a link by SMS.

  • content: Include a short message and ask for consent.
  • actions: If accepted, call start_sms_flow.

Example – Golf info

​
Conditional

Let the user choose which SMS to receive.

  • content: Ask a branching question.
  • actions: Map answers to different sms_ids.

Example – Tee time booking

​
Info only

Used for reference material (like opening hours or prices). No interaction or action required.

  • Fill content
  • Leave actions empty

Example – Opening hours

​
Best practices

Do this:

  • Use specific topic names
  • Add 5–10 realistic sample phrasings
  • Keep replies short, helpful, and on-brand
  • Use actions only when necessary
  • Split multi-part flows into separate turns or topics

Avoid this:

  • Bundling multiple intents into one topic
  • Running more than one action per turn
  • Using vague topic names like Misc, Help, or Info

​
Resources

The knowledge base (KB) is the brain of your agent. Each topic defines a single intent, how users might phrase it, what the agent should say, and what it should do next.

Well-structured entries keep answers clear, safe, and meaningful.

​
How retrieval works

The agent does not see the whole knowledge base every time. Instead, it uses retrieval-augmented generation (RAG) to find the best match for the user’s message.

  1. The retriever compares the user’s message to the name, sample questions, and content.
  2. It returns the closest few topics.
  3. The model selects one and generates a reply (and may trigger an action).

​
Types of knowledge base entries

We will now go through the broad categories that most knowledge base entries tend to fall into:

​
Simple FAQ

Used when the agent just needs to answer a question with no follow-up or action required.

​
Single turn

One message in, one message out.

  • Add a short, helpful reply in content.
  • Leave actions empty.

Example – Pet policy (single-turn)

​
Multi turn

Ask a clarifying question before giving the answer.

  • Use branching inside content to structure the reply.
  • Leave actions empty.

Example – Pet policy (multi-turn)

​
Handoff

Used when the agent should offer to connect the user to a human agent.

​
Offer

Agent replies, then offers to transfer.

  • content: Include the answer + offer.
  • actions: Trigger transfer_call only if user accepts.

Example – Spa booking

​
Direct

Immediate transfer with no agent reply.

  • Leave content empty.
  • Always run transfer_call in actions.

Example – Billing handoff

​
Conditional

Transfer based on user clarification, for example, group size, request type.

  • content: Ask a disambiguating question.
  • actions: Map each answer to the correct destination.

Example – Room reservations

​
SMS

Used when you want to follow up with a link or info through text.

​
Offer

Offer to send a link by SMS.

  • content: Include a short message and ask for consent.
  • actions: If accepted, call start_sms_flow.

Example – Golf info

​
Conditional

Let the user choose which SMS to receive.

  • content: Ask a branching question.
  • actions: Map answers to different sms_ids.

Example – Tee time booking

​
Info only

Used for reference material (like opening hours or prices). No interaction or action required.

  • Fill content
  • Leave actions empty

Example – Opening hours

​
Best practices

Do this:

  • Use specific topic names
  • Add 5–10 realistic sample phrasings
  • Keep replies short, helpful, and on-brand
  • Use actions only when necessary
  • Split multi-part flows into separate turns or topics

Avoid this:

  • Bundling multiple intents into one topic
  • Running more than one action per turn
  • Using vague topic names like Misc, Help, or Info

​
Resources