Skip to main content

Customizing Patterns

Rasa assistants use patterns (pre-built flow templates) to handle conversation repair and some system behaviour. By separating patterns from the business-logic flows, you avoid having to weave every possible “what if the user changes their mind” scenario into your main flows. Instead, your flows can stay focused on the high-level steps for accomplishing tasks (booking a flight, transferring money, etc.), while the conversation patterns makes sure your flows are adaptive to how the user wants to accomplish their goal.

By default, every common conversation repair scenario—like corrections, digressions, or incomplete data—has a pre-written pattern in Rasa Pro. However, you can override those defaults to better suit your assistant’s needs.

Digging into patterns

For more information on the concept of patterns, head to the Home section on Patterns.

For a complete list of patterns and their low-level implementations, head to the corresponding section in the Reference.

Why Customise Patterns?​

Templated Flows for Repair​

Think of a pattern as a reusable template for handling a repeatable scenario (cancellation, correction, human handoff, etc.). Because patterns need to work anywhere in your assistant, you typically want to keep them general rather than making them skill-specific.

Balancing Consistency and Adaptation​

Although you want to keep these templated flows broadly applicable, you may still want to:

  • Adapt the language of responses (e.g., brand or domain-specific wording).
  • Request user confirmation before updating slots (such as in the correction pattern).
  • Hand over to a live agent when certain issues arise.

For purely language-level customization (i.e., to keep brand voice consistent), consider using the Response Rephraser to automatically adapt the text in the pattern’s default responses. You can find more information on customizing the tone of your assistant here.

How to Customise Patterns​

1. Override a Pattern Flow​

To customise a pattern, create a flow in your flows.yml file with the same name as the pattern. For instance, if you want to change how a cancellation is handled, create a flow named pattern_cancel_flow:

flows.yml
flows:
  pattern_cancel_flow:
    description: Custom cancellation flow
    steps:
      - action: action_cancel_flow
      - action: utter_flow_cancelled_rasa

When your assistant sees a user cancellation, it will use your new pattern_cancel_flow instead of the default one.

For existing default pattern implementations see to the Reference section on Patterns.

Keep pattern overrides concise. Most patterns interrupt an ongoing flow, so it’s best to return the user to the main conversation quickly.

2. Override Default Actions or Responses​

Some patterns make use of default actions, like action_correct_flow_slot or action_trigger_chitchat. If you need more control over the logic:

  1. Create a custom action in your actions.py.
  2. Reference it in your custom pattern flow and your domain.yml.

Likewise, the default text responses are in your domain’s responses: section. Override them with your own text or tie them into the Rephraser if needed.

You can also “jump” from a pattern to another flow (for example, if the user clarifies something and you want to start a brand-new flow). To do this, add a link step at the end of your pattern:

flows.yml
flows:
  pattern_clarification:
    description: "Clarification pattern"
    steps:
      - action: action_clarify_flows
        next:
          - if: context.names contains "transfer_money"
            then:
              - link: transfer_money

When the user clarifies that they want to transfer money, you directly start the transfer_money flow.

4. Human Handoff and Other Special Cases​

For certain special patterns like pattern_human_handoff or pattern_chitchat:

  • Human handoff: Override pattern_human_handoff to add your own logic for connecting to a live agent.
  • Chitchat: By default, chitchat uses a short pre-written response. You can switch to free-form generation by overriding pattern_chitchat and calling an LLM-based response. Or if you’d rather disable chitchat, you can override it to send a single “I can’t handle that” response.

For more information on customizing patterns, see to the Reference section on Patterns.

Last updated on