Ways of Customizing Assistant Tone in Rasa

There are two main ways to adapt your assistant’s messaging style:

1. Response Rephraser
   Dynamically rewrites your responses using an LLM prompt. You can instruct the LLM to adopt or maintain a specific personality or tone.

2. Conditional Response Variations
   Predefine different text variations (often for different slot conditions, contexts, or channels). This method is more static, but ensures each user scenario has a hand-crafted, on-brand response.

What Is the Response Rephraser?

The Response Rephraser is an LLM-powered component that takes a templated response (for example, “I’m sorry, I can’t help with that”) and rephrases it while preserving your original meaning and factual content. Its key benefits are:

• Dynamically adapting tone: Use a single base response but produce many variations, all preserving your brand guidelines.
• Contextual awareness: The rephraser can read the conversation history and user input, ensuring rephrasings make sense in context.
• Easy maintenance: If you decide to tweak your brand style (e.g., become more informal), you can simply update the LLM prompt, rather than rewriting all your responses.

What Are Conditional Response Variations?

Conditional response variations let you define several static message templates under the same response name but tailor them to different states of the conversation. For example, you might have:

• A casual greeting if the user_is_repeat_customer slot is set to true.
• A more formal greeting otherwise.

When the assistant triggers the response, CALM checks any conditions you’ve set (slot values, channel, etc.) and picks the matching variant. These variations are not LLM-based—they are “what you see is what you get” messages, drawn from your domain or responses files.

When to Use Which?

• Use the Response Rephraser if you want:
  • Dynamic, LLM-powered rewording that can adapt to your brand voice, summarizing or refining the text while retaining the original meaning.
  • A simpler way to unify tone across your entire assistant, especially for messages that appear in repair patterns or emergent flows.
• Use Conditional Response Variations if you want:
  • Strict control over the exact wording in specific contexts or channels (e.g., “VIP members get a special greeting”).
  • Variation without an external LLM call, or you need a guaranteed brand-approved statement for certain user segments.

Of course, you can also combine them—some responses can have multiple static variants, and you still run them through the Rephraser for final polishing.

How to Create Conditional Response Variations

1. In your domain file, provide multiple responses under the same response name.
2. Add a condition block for each variant to specify which slot values must match (or which channel must match).
3. Always include a default fallback response (with no condition) in case none of the conditions are met.

responses:
  utter_greet:
    - condition:
        - type: slot
          name: logged_in
          value: true
      text: "Hey, welcome back! How are you?"
    - text: "Hello! How can I help you today?"

How to Customize the Response Rephraser

You can configure the Response Rephraser to ensure it outputs messages aligned with your desired personality or brand identity.

1. Enabling Rephraser Across All Responses

In your endpoints.yml file, add:

nlg:
  type: rephrase
  rephrase_all: true

This automatically attempts to rephrase every utterance. If you want some responses left untouched, annotate them with metadata: { rephrase: false } in your domain.

2. Enabling Rephraser for Specific Responses

If you prefer a more selective approach:

• Enable type: rephrase in endpoints.yml without setting rephrase_all: true.
• Add metadata: { rephrase: true } to only the responses you’d like rephrased:

domain.yml

responses:
    utter_greet:
    - text: "Hello there!"
      metadata:
        rephrase: true

3. Setting a Custom Prompt

You can supply your own Jinja2 prompt template to the rephraser. This is especially important to define a tone or style. For example:

endpoints.yml

nlg:
  type: rephrase
  prompt: "prompts/brand-tone-rephraser-template.jinja2"

Inside that .jinja2 file, you could add instructions such as:

“Use a casual, friendly tone in second-person. Always address the user by name if available.”

You can also override the default prompt for a single response by setting rephrase_prompt in its metadata (see the reference docs for an example).

How to Test Rephrased Responses?

Because the final assistant message may be partially (or entirely) generated by an LLM, it’s crucial to test for both correctness and style:

1. Generative Response Is Relevant Ensures the rephrased message is on-topic and aligns with the user’s query.

tests/e2e_test_cases.yml

assertions:
  - generative_response_is_relevant:
      threshold: 0.90

2. Generative Response Is Grounded Ensures the rephrased message remains factually accurate to the original domain response or RAG context.

tests/e2e_test_cases.yml

assertions:
  - generative_response_is_grounded:
      threshold: 0.90
      ground_truth: "Free upgrades require a Platinum membership."

You can add these assertions in end-to-end tests to confirm that the LLM’s rewriting (1) remains brand-appropriate, (2) is accurate, and (3) is still relevant to the user’s input. See E2E Testing for more on these assertion types.