Rasa Rephrasing Responses

Rephrasing Responses in Rasa

The Contextual Response Rephraser uses an LLM to dynamically rewrite templated responses, making them sound more natural and context-aware while preserving the original meaning. It reads conversation history and user input to ensure rephrasings fit the context.

Workflow

1. Enable the rephraser in endpoints.yml (see "Endpoints configuration").
2. Decide the rephrasing scope — per-response or all responses

(see "Rephrasing scope").

1. Add metadata: rephrase: True/False on individual responses as needed

(see "Domain-side metadata").

1. Optionally customize the prompt globally or per response

(see "Prompt customization").

Endpoints configuration

Enable the rephraser by adding nlg: type: rephrase to endpoints.yml.

Configure the LLM model and temperature. Lower temperature (default 0.3) produces more predictable rephrasings; higher temperature produces more variable output but risks altering meaning.

nlg:
  type: rephrase
  llm:
    model_group: my_llm               # optional, defaults to the default model

model_groups:
  - id: my_llm
    models:
      - provider:      # e.g. openai, azure, self-hosted
        model: 
        temperature: 0.3              # 0.0–2.0, default 0.3

Conversation history

Two modes for how conversation history is included in the rephrasing prompt.

Summary mode (default) — summarizes history using an additional LLM call. No extra config needed.

Transcript mode — keeps the last n turns as-is. Set summarize_history: False and optionally adjust max_historical_turns (default 5).

nlg:
  type: rephrase
  summarize_history: False
  max_historical_turns: 5

Rephrasing scope

Specific responses only (default)

No endpoints.yml change needed. Add metadata: rephrase: True on each response you want rephrased.

responses:
  utter_greet:
    - text: "Hey! How can I help you?"
      metadata:
        rephrase: True

All responses

Set rephrase_all: true in endpoints.yml. Every response is rephrased without needing per-response metadata.

nlg:
  type: rephrase
  rephrase_all: true

All except specific ones

Set rephrase_all: true in endpoints.yml, then opt out individual responses with metadata: rephrase: False.

responses:
  utter_legal_disclaimer:
    - text: "By proceeding you agree to our terms and conditions."
      metadata:
        rephrase: False

Prompt customization

Reference a custom Jinja2 template globally in endpoints.yml, or override per response via rephrase_prompt in metadata.

Available variables: {{history}}, {{current_input}}, {{suggested_response}}.

# Global prompt — endpoints.yml
nlg:
  type: rephrase
  prompt: prompts/response-rephraser-template.jinja2

# Per-response prompt — domain.yml
responses:
  utter_greet:
    - text: "Hey! How can I help you?"
      metadata:
        rephrase: True
        rephrase_prompt: |
          Rephrase the suggested response in a friendly, casual tone.
          Stay close to the original meaning.
          Context: {{history}}
          {{current_input}}
          Suggested: {{suggested_response}}
          Rephrased:

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: RasaHQ
• Source: RasaHQ/rasa-agent-skills
• License: Apache-2.0

Install and usage instructions live in the source repository linked above.