Rasa Building Flows

Building Flows in Rasa

Workflow

1. Identify the business process the user wants to implement.
2. Review existing flows, slots, responses, and actions. Reuse before creating new ones.
3. Scope the work — one file per business process, one flow per user goal (see "Scoping a flow").
4. Create or open the YAML file under data/, named after the primary flow (see "File organization").
5. Define the flow with a snake_case ID (see "Naming conventions").
6. Write the flow description (see "Writing descriptions").
7. Add steps: collect for user input, action for responses or invoking custom actions.
8. Write descriptions for collect steps (see "Writing descriptions").
9. Add branching with next/if/then/else where the process has decision points (see "Branching and conditions").
10. Extract reusable sub-tasks into child flows via call (see "Connecting flows").
11. Connect follow-up flows via link if a separate process should start after.
12. Ensure all referenced slots, responses, and actions exist in the domain (see "Ensure domain completeness").
13. Validate the project (see "Fixing validation errors").

Scoping a flow

A flow represents one business process from the user's perspective. Use these questions to decide whether something warrants its own flow:

1. Does a user explicitly ask for this? → yes = flow
2. Can it run independently from start to finish? → yes = flow
3. Is it a sub-task reused by multiple flows? → yes = child flow with if: False
4. Is it just one action at the end of another flow? → no, keep it inline

Right-sized flow - a single user goal with a clear start and end:

flows:
  track_order:
    description: Lets users check the delivery status of an existing order.
    steps:
      - collect: order_id
      - action: action_fetch_order_status
      - action: utter_order_status
      # ... can also include branching, call, link steps as needed

Too big - multiple unrelated goals in one flow. Split them:

# BAD
flows:
  shopping:
    description: Handles all shopping operations.
    steps:
      - collect: operation_type
      # ... 20 or more steps mixing order tracking, returns, payments,
      # branching, sub-flow calls — too many unrelated concerns in one place

Too small - a single step that is not a business process. Inline it:

# BAD
flows:
  say_goodbye:
    description: Says goodbye.
    steps:
      - action: utter_goodbye

File organization

Before creating new files, check the existing project layout. If the project already uses a different convention, follow that convention instead.

By default, all flow files live under data/. Subdirectories are supported — Rasa recursively scans data/ for YAML files. Each file groups a business process with its sub-flows:

data/
├── book_appointment.yml        # flow + child flows in one file
├── track_order.yml
├── transfer_money.yml
└── onboarding/                 # subdirectory for a larger domain area
    ├── register_patient.yml
    └── verify_insurance.yml

Name the file after the primary flow it contains. If book_trip has child flows like select_destination and choose_flights, all three go in data/book_trip.yml.

Use subdirectories to group related flows by domain area when the project grows large (e.g. data/payments/, data/scheduling/). The directory structure has no functional impact — it's purely organizational.

Naming conventions

This section proposes default conventions. If the user's project already follows different conventions, match those instead.

Flow IDs

• Use snake_case: transfer_money, book_appointment, track_order
• Start with a verb: check_balance, schedule_visit — not flow_1 or appointment
• Child flows can use a prefix of the parent for clarity: book_appointment_collect_patient_info

Flow names

The optional name field is a human-readable label. It has no functional impact.

flows:
  book_appointment:
    name: "Book Appointment"
    description: Helps patients schedule a visit with their doctor.

Writing descriptions

Flow descriptions

The description is the primary signal the LLM within the Command Generator uses to decide when to start a flow. A bad description means the flow won't trigger — or triggers for the wrong request.

• Be information-dense: precisely outline the flow's purpose and scope. Avoid filler words. Use imperative language.
• Use clear, standard language: avoid unusual phrasing. Stick to universally understood terms.
• Explicitly define context: the embedding model lacks situational awareness - it can't read between the lines. State the domain and intent directly.
• Clarify specialized knowledge: if the flow references brand names or product names, explain what they are. The embedding model won't know.
• Keep descriptions distinct: if two flows sound similar, add distinguishing details so the LLM within the Command Generator can disambiguate.

# GOOD — information-dense, user perspective, explicit context
book_appointment:
  description: Helps patients schedule a visit with their doctor.

track_order:
  description: Lets users check the delivery status of an existing order.

# BAD — implementation detail, vague, no context
book_appointment:
  description: Runs action_book_slot and fills the date slot.

track_order:
  description: A flow for orders.

Collect step descriptions

The description on a collect step tells the LLM within the Command Generator what value to extract from the user's message. Without it, the LLM within the Command Generator only has the slot name to guess.

• Describe the expected value, not the question being asked
• Include format, constraints, or valid values when applicable
• Provide examples for ambiguous or structured data

# GOOD
- collect: insurance_id
  description: >
    Insurance member ID printed on the patient's card.
    Alphanumeric, 9-12 characters (e.g. XYZ-123456789).

- collect: return_reason
  description: "Reason for the return: damaged, wrong_item, changed_mind, or other."

# BAD — no useful information / describes the question, not the value
- collect: amount
  description: the amount
- collect: email
  description: Ask the user for their email address.

Slots and collect steps

By default, slots filled in collect or set_slots steps reset when the flow ends (to null or to the slot's initial_value from the domain). To retain a slot's value after the flow completes, list it in persisted_slots at the flow level. Slots set in custom actions are persistent by default and should not be listed in persisted_slots.

flows:
  return_item:
    description: Helps users return a purchased item for a refund or exchange.
    persisted_slots:
      - order_id         # kept after flow ends
      - return_reason    # kept after flow ends
    steps:
      - collect: order_id
      - collect: return_reason
      - action: action_process_return

A collect step automatically uses a response named utter_ask_{slot_name} to prompt the user. Override this with the utter property on the collect step, or use a custom action named action_ask_{slot_name} instead — but not both.

- collect: account_type
  utter: utter_ask_secondary_account_type   # overrides default utter_ask_account_type

Use rejections on a collect step to validate values inline. The assistant re-asks automatically on rejection. For complex validation, create a custom action named validate_{slot_name}.

- collect: age
  rejections:
    - if: slots.age < 1
      utter: utter_invalid_age
    - if: slots.age < 18
      utter: utter_must_be_18

To reset a slot mid-flow, set it to null in a set_slots step. In a custom action, return None. Empty slots are not eligible for user correction.

- set_slots:
    - amount: null

See the full reference for all collect step properties (ask_before_filling, force_slot_filling, utter, etc.).

Branching and conditions

Any step can branch by adding a next property with if/then/else conditions. Conditions support operators like and, or, not, <, >, =, !=, is, is not, contains, and matches (see full reference for details).

Two namespaces are available in conditions:

• slots. — slot values: slots.age < 18, slots.plan = "premium"
• context. — dialogue frame properties: context.previous_flow_name, context.collect, context.canceled_name

When context.collect or context.validate reference a slot name, wrap them in Jinja with the slots. prefix: "slots.{{context.collect}} is not null"

When a condition leads to next: END, the flow completes without running further steps. Use a noop step when you need a branch point without performing any action.

- collect: membership_tier
  next:
    - if: slots.membership_tier = "premium"
      then: premium_step
    - if: slots.membership_tier = "standard"
      then: standard_step
    - else:
        - action: utter_unknown_tier
          next: END
- id: premium_step
  action: utter_premium_benefits
- id: standard_step
  noop: true
  next:
    - if: not slots.verified_email
      then:
        - call: verify_email
          next: continue_step
    - else: continue_step
- id: continue_step
  action: utter_standard_benefits

Connecting flows: call vs link

Two step types connect flows together. Choose based on whether the current flow should continue or end:

Example — an orchestrator flow that sequences call steps for sub-tasks and ends with link to hand off to a separate process. Child flows should not overlap in the slots they collect, so the LLM within the Command Generator can unambiguously route user input.

flows:
  book_trip:
    description: Guides the user through booking a complete trip with flight and hotel.
    steps:
      - call: book_flight                  # sub-task: returns here when done
      - call: book_hotel                   # sub-task: returns here when done
      - action: utter_trip_summary
      - link: process_payment              # done — hand off to a separate process

  book_flight:
    description: Helps the user search and select a flight.
    if: False
    steps:
      # ... collect destination, dates, search, select flight

  book_hotel:
    description: Helps the user search and select a hotel at the destination.
    if: False
    steps:
      # ... search hotels, select hotel

  process_payment:
    description: Handles payment for a completed booking using a saved or new card.
    steps:
      # ... independent business process, starts fresh after link

Flow guards

Guards control whether a flow can be started by the LLM within the Command Generator:

# Only start if patient is verified
view_medical_records:
  if: slots.authenticated AND slots.identity_verified

# Never start directly — only via call or link
collect_patient_info:
  if: False

Ensure domain completeness

Every element a flow references must be defined in the domain (domain.yml or split domain files). After writing or editing a flow, walk through each step and verify the items below are present. Missing entries cause validation errors or silent runtime failures.

• Every slot used in the flow → slots: (see rasa-managing-slots)
• Every collect: slot_name → utter_ask_{slot_name} in responses: (see rasa-writing-responses)
• Every action: utter_* step → matching entry in responses:
• Every custom action / validation action → actions: (see rasa-writing-custom-actions)
• Rejection responses → responses:

Fixing validation errors

When validation fails, preserve the user's intent — what the flow is supposed to accomplish. The flow steps themselves can change, but the intended behavior must not be lost.

Never strip down a flow's intended behavior to make validation pass. If needed, fix the domain: slots, responses, and actions to support what the user asked for.