Skip to main content

Overview

You can customise many aspects of how your assistant project works by modifying the following files: config.yml, endpoints.yml, and domain.yml.

A minimal configuration for a CALM assistant looks like this:

config.yml
recipe: default.v1
language: en
assistant_id: 20230405-114328-tranquil-mustard

pipeline:
  - name: CompactLLMCommandGenerator

policies:
  - name: rasa.core.policies.flow_policy.FlowPolicy
Default Configuration

For backwards compatibility, running rasa init will create an NLU-based assistant. To create a CALM assistant with the right config.yml, add the additional --template argument:

rasa init --template calm

Assistant ID​

The assistant_id key should be a unique value and allows you to distinguish multiple deployed assistants. This id is added to each event's metadata, together with the model id. See event brokers for more information. Note that if the config file does not include this required key or the placeholder default value is not replaced, a random assistant name will be generated and added to the configuration every time you run rasa train.

Recipe​

The recipe key only needs to be modified if you want to use a custom graph recipe. The vast majority of projects should use the default value "default.v1".

Language​

  • The language key sets the primary language your assistant supports. Use a two-letter ISO 639-1 code (e.g., "en" for English).
  • additional_languages key lists codes of other languages your assistant supports.

With these settings, your assistant will default to its primary language but can recognize and respond in all configured languages. You can further translate your assistant’s content. For more details, refer to our Translating Your Assistant guide.

Here’s the example for assistant which is using English as default while also supporting Italian, German, and French:

config.yml

# ...
language: "en"# Default language: English
additional_languages:
  - "it"# Italian
  - "de"# German
  - "fr"# French
# ...

You can use any valid language or locale-specific code following the BCP 47 standard:

  • Basic language codes: e.g., "en", "de", "it".
  • Locale-specific codes: e.g., "en-US", "fr-CA", "de-CH".
  • Custom language codes: e.g., "x-en-formal".

Make sure all language codes adhere strictly to this format to avoid unexpected validation errors.

BCP 47 Standard

Rasa adheres to the BCP 47 standard for language codes. This ensures compatibility with platforms such as Twilio Voice, Genesys Cloud, and Amazon Connect.

Pipeline​

The pipeline key lists the components which will be used to process and understand the messages that end users send to your assistant. In a CALM assistant, the output of your components pipeline is a list of commands.

The main component in your pipeline is the LLMCommandGenerator. Here is what an example configuration looks like:

config.yml
  pipeline:
    - name: CompactLLMCommandGenerator
      llm:
        model_group: openai_llm
      flow_retrieval:
        embeddings:
          model_group: openai_embeddings
      user_input:
        max_characters: 420
endpoints.yml
   model_groups:
     - id: openai_direct
       models:
         - model: "gpt-4o-2024-11-20"
           provider: "openai"
           timeout: 7
           temperature: 0.0
     - id: openai_embeddings
       models:
         - model: "text-embedding-3-large"
           provider: "openai"

The full set of configurable parameters is listed here.

All components which make use of LLMs have common configuration parameters which are listed here

Policies​

The policies key lists the dialogue policies your assistant will use to progress the conversation.

config.yml
policies:
  - name: rasa.core.policies.flow_policy.FlowPolicy

The FlowPolicy currently doesn't have an additional configuration parameters.

Silence Timeout Handling​

Silence timeouts help your assistant handle situations where the user doesn’t respond. For now, this setting only works with voice-stream channels, such as:

  • Twilio Media Streams
  • Browser Audio
  • Genesys
  • Jambonz Stream
  • Audiocodes Stream

There are two types of timeouts you can configure.

Global Silence Timeout​

You can set a default silence timeout across your assistant by adding this to your endpoints.yml:

interaction_handling:
  global_silence_timeout: 7

This means the assistant will wait 7 seconds (or your configured value) for a user reply before treating it as silence and triggering fallback logic.

Local (Per-Step) Silence Timeout​

You can override the global value for specific Collect steps:

steps:
  - collect:
      name: ask_email
      silence_timeout: 10

This allows you to fine-tune the timing for specific questions. For example, you may want to:

  • Wait longer on more complex or sensitive questions (e.g., "Can you describe your issue?")
  • Use shorter timeouts for quick prompts (e.g., yes/no questions)

Tailoring silence handling this way improves the conversational experience.

Enabling/Disabling Silence Timeouts​

If you want to disable silence detection so it never triggers during a conversation, you can set the timeout to a very high value. For example, to disable it globally:

interaction_handling:
  global_silence_timeout: 70000

Use this approach if you want to avoid fallback interruptions but still need a valid numeric value for configuration or platform compatibility.

Customizing the Assistant's Response to Silence​

When the silence timeout is reached, the assistant triggers the pattern_user_silence. You can customize how your assistant responds to silence by modifying this pattern.

👉 Learn more about patterns configuration

Last updated on