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