Rasa Writing E2e Tests

Writing E2E Tests in Rasa

E2E tests validate the assistant as a whole system — from user message to bot response, including flow execution, slot filling, custom actions, and generative responses. They live as YAML files in the tests/ directory and run via rasa test e2e.

Workflow

1. Identify the conversation scenario to test (happy path, edge case, error path).
2. Review existing tests — extend before creating new ones.
3. Create or open a YAML file under tests/ (see "File organization").
4. Write the test case with user steps and expected outcomes

(see "Writing test cases").

1. Add assertions for detailed validation if needed (see "Assertions").
2. Set up fixtures if the test requires pre-filled slots (see "Fixtures").
3. Stub custom actions if needed to avoid external calls (see "Stubbing custom actions").
4. Run tests with rasa test e2e (see "Running tests").

Key principles

• One test case per scenario: happy path, cancellation, validation failure, etc.
• Start with a user step: the test runner begins evaluation after

action_session_start.

• Alternate user and expected outcomes: avoid consecutive user steps.
• Use assertions for CALM-specific checks: flow_started, flow_completed,

action_executed give stronger guarantees than matching bot text.

• Use bot or utter steps for simple cases: when you only need to verify the

response text or domain response name.

• Use fixtures to avoid repetitive setup: shared context like authentication goes

in conftest.yml.

File organization

Test files live under tests/ and are automatically discovered by Rasa. Subdirectories are supported — pass the path as a positional argument to rasa test e2e.

tests/
├── e2e_test_cases.yml           # default test file
├── conftest.yml                 # shared fixtures (visible to all tests)
├── booking/
│   ├── conftest.yml             # fixtures scoped to this directory
│   ├── test_book_trip.yml
│   └── test_cancel_booking.yml
└── payments/
    └── test_process_payment.yml

Each file must contain the test_cases key with a list of test cases. Each test case needs a unique test_case name and a list of steps.

Writing test cases

A test case simulates a conversation turn by turn. After each user step, the test runner captures events from the tracker and compares them with expected steps. Avoid multiple consecutive user steps without bot responses between them — the runner only evaluates events from the most recent user step.

Step types

| Step | Purpose | |--------------------|---------| | user | Simulates a user message | | bot | Checks exact text of bot response | | utter | Checks domain response name (e.g. utter_ask_email) | | slot_was_set | Confirms a slot was set, optionally with a value | | slot_was_not_set | Confirms a slot was not set or has a different value |

The order of bot, utter, and slot steps after a user step does not matter — they just need to occur after that user turn. Extra events beyond those specified are ignored and do not cause failures.

test_cases:
  - test_case: user schedules an appointment
    steps:
      - user: I need to see Dr. Smith next Tuesday
      - slot_was_set:
          - doctor_name: "Dr. Smith"
          - preferred_date                     # checks slot is set, any value
      - utter: utter_ask_appointment_time
      - user: "10am"
      - slot_was_set:
          - appointment_time: "10am"
      - utter: utter_confirm_appointment
      - user: "yes"
      - bot: "Your appointment with Dr. Smith has been confirmed."

Assertions

For more detailed checks, attach assertions to a user step. When a step has assertions, the basic step types (bot, utter, slot_was_set) are ignored for that step — use the corresponding assertion types instead.

| Assertion | Purpose | |-----------------------------------|---------| | flow_started | Verify a flow began | | flow_completed | Confirm a flow (and optionally a step) finished | | flow_cancelled | Ensure a flow was cancelled | | pattern_clarification_contains | Check clarification suggestions returned | | slot_was_set | Validate slot name and value | | slot_was_not_set | Confirm slot is not set or has a different value | | action_executed | Assert an action was triggered | | bot_uttered | Verify response text (regex), buttons, or utter name | | bot_did_not_utter | Ensure bot did not respond with specific content | | generative_response_is_relevant | Check generative response relevance (threshold 0–1) | | generative_response_is_grounded | Check factual accuracy against ground truth (threshold 0–1) |

test_cases:
  - test_case: user tracks an order
    steps:
      - user: "Where is my order?"
        assertions:
          - flow_started:
              operator: "all"
              flow_ids:
                - "track_order"
          - bot_uttered:
              utter_name: utter_ask_order_id
      - user: "ORD-12345"
        assertions:
          - slot_was_set:
              - name: "order_id"
                value: "ORD-12345"
          - action_executed: "action_fetch_order_status"
          - bot_uttered:
              text_matches: "Your order .* is currently .*"  # regex

Generative response assertions

Use generative_response_is_relevant and generative_response_is_grounded to test outputs from the Contextual Response Rephraser, Enterprise Search, or custom actions. Specify utter_source to target a specific component.

- user: "What is your return policy?"
  assertions:
    - generative_response_is_relevant:
        threshold: 0.85
        utter_source: EnterpriseSearchPolicy
    - generative_response_is_grounded:
        threshold: 0.90
        utter_source: EnterpriseSearchPolicy
        ground_truth: "Items can be returned within 30 days of purchase."

The LLM Judge model is configured in conftest.yml at the project root (if not specified, Rasa falls back to its default model):

llm_judge:
  llm:
    provider:            # e.g. openai, azure, self-hosted
    model: 
  embeddings:
    provider: 
    model: 

Fixtures

Fixtures pre-fill slots before a test case runs — useful for testing scenarios that require context (e.g. logged-in user, specific membership tier). Define fixtures at the top of a test file or in conftest.yml for shared access.

Fixture override order: root conftest < folder conftest < test file. Slots are set after action_session_start and before the first step.

fixtures:
  - verified_user:
      - is_authenticated: true
      - membership_tier: "premium"

test_cases:
  - test_case: premium user books a trip
    fixtures:
      - verified_user
    steps:
      - user: "I want to book a trip"
      - utter: utter_ask_destination

Mocking datetime

For deterministic testing of time-dependent behavior, use the built-in mocked_datetime slot in fixtures (ISO 8601 format):

fixtures:
  - fixed_date:
      - mocked_datetime: "2025-06-15T10:00:00+00:00"

test_cases:
  - test_case: user asks about today
    fixtures:
      - fixed_date
    steps:
      - user: "What day is it?"
      - bot: "Today is Sunday, June 15, 2025."

Stubbing custom actions

Stub custom actions to test without running the action server. Define stub_custom_actions at the top level of the test file. When stubbing is used, all custom actions called in the file must be stubbed.

Requires RASA_PRO_BETA_STUB_CUSTOM_ACTION=true.

stub_custom_actions:
  action_fetch_order_status:                   # applies to all test cases in this file
    events:
      - event: slot
        name: order_status
        value: "shipped"
    responses:
      - text: "Your order has been shipped."

  test_express_delivery::action_fetch_order_status:   # scoped to one test case
    events:
      - event: slot
        name: order_status
        value: "out_for_delivery"
    responses:
      - text: "Your order is out for delivery."

Running tests

rasa test e2e                                  # run all tests in tests/
rasa test e2e tests/booking/                   # run tests in a subdirectory
rasa test e2e --fail-fast                      # stop at first failure
rasa test e2e --coverage-report                # generate flow coverage report

If custom actions are not stubbed, start the action server first:

rasa run actions &
rasa test e2e

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.