Rasa Configuring Assistant

Configuring a Rasa Assistant

Workflow

1. Check if config.yml and endpoints.yml already exist.
2. Configure the pipeline in config.yml with at least one command generator (see "Pipeline").
3. Add FlowPolicy to policies (see "Policies").
4. Set the language and assistant ID (see "Language" and "Assistant ID").
5. Define model groups in endpoints.yml.
6. Configure the action endpoint if custom actions are used (see "Action endpoint").
7. Validate the project.

config.yml

The config.yml file defines how the assistant processes user messages. It specifies recipe, language, pipeline components, and policies.

Minimal CALM configuration

A working CALM assistant requires at minimum:

recipe: default.v1
language: en

pipeline:
  - name: CompactLLMCommandGenerator  # or SearchReadyLLMCommandGenerator

policies:
  - name: FlowPolicy

Pipeline

The pipeline processes user messages and produces commands for the conversation. The main component is the Command Generator (e.g. CompactLLMCommandGenerator or SearchReadyLLMCommandGenerator), which uses an LLM to interpret user messages and generate commands.

Configure the LLM model via model_group (defined in endpoints.yml), flow retrieval embeddings, and input limits:

pipeline:
  - name: CompactLLMCommandGenerator
    llm:
      model_group: my_llm                  # references model_groups in endpoints.yml
    flow_retrieval:
      embeddings:
        model_group: my_embeddings         # references model_groups in endpoints.yml
    user_input:
      max_characters: 420

Policies

Policies determine how the assistant progresses conversations. For CALM, FlowPolicy is required — it executes flow steps based on the commands produced by the pipeline. No additional configuration is needed.

policies:
  - name: FlowPolicy

Language

Set the primary language with a two-letter ISO 639-1 code. Use additional_languages for multilingual assistants.

language: en
additional_languages:
  - de
  - fr

Assistant ID

A unique identifier included in every event's metadata. Always set this explicitly — if missing, a random ID is generated on every rasa train.

assistant_id: my_assistant

endpoints.yml

The endpoints.yml file defines how the assistant connects to external services — LLM providers, action servers, model storage, and more.

Use ${VARIABLE_NAME} to reference environment variables for API keys and other sensitive values.

Model groups

Define model groups in endpoints.yml for LLM and embedding providers. Pipeline components reference groups by id.

See the rasa-configuring-model-groups skill for full details on providers, multi-deployment routing, failover, and self-hosted models.

model_groups:
  - id: my_llm
    models:
      - provider: <your-provider>       # e.g. openai, azure, self-hosted
        model: <your-llm-model>

  - id: my_embeddings
    models:
      - provider: <your-provider>
        model: <your-embedding-model>

Action endpoint

Tells Rasa where the action server runs for executing custom actions. Supports HTTP, HTTPS, gRPC, and secure gRPC protocols.

Use enable_selective_domain: true to only send the domain to actions that explicitly request it (reduces payload size).

action_endpoint:
  url: "http://localhost:5055/webhook"   # or https://, grpc://
  # cafile: "/path/to/ssl_ca_certificate"  # for HTTPS or secure gRPC
  enable_selective_domain: true            # optional, reduces payload size

NLG server

Configure an external NLG server to generate responses dynamically instead of using static templates from the domain. The endpoint must serve a /nlg path.

nlg:
  url: http://localhost:5055/nlg
  # token: "my_authentication_token"        # optional token auth
  # basic_auth:                             # or basic auth
  #   username: user
  #   password: pass

The rephraser (nlg: type: rephrase) is covered by the rasa-rephrasing-responses skill.

MCP servers

MCP server configuration (mcp_servers in endpoints.yml) is covered by the rasa-configuring-mcp-server skill.

Silence handling

Controls how long the assistant waits before assuming the user is silent. Only applies to voice-stream channels (Twilio, Browser Audio, Genesys, Jambonz, Audiocodes).

interaction_handling:
  global_silence_timeout: 7    # seconds, default: 7