Building Assistive Conversational Agents with Rasa Pro (CALM)

Fred Agbo

2026-10-21

👋 Welcome to Lab 1

Course: Human-AI Interaction — Willamette University
Format: Hands-on lab (CLI/IDE workflow, no Rasa Studio required)
Outcome: Build, train, test, and locally deploy two working assistants

🎯 Lab Overview

By the end of lab, you will:

1. Install Rasa Pro and set up Python environment
2. Initialize the CALM tutorial starter project
3. Build Scenario 1: Course FAQ Assistant
4. Build Scenario 2: IT Help-Desk Ticket Assistant
5. Diagnose and fix common YAML and flow errors

🧠 Why CLI/IDE Instead of Studio?

• Makes flows, slots, triggers, and branching explicit as text
• Better for learning how CALM actually works under the hood
• Strong foundation before moving to visual/no-code tooling

In this lab, we optimize for conceptual understanding, not just click-through completion.

🧰 Part 0 — Prerequisites & Installation

0.1 Check Python version (3.10–3.13)

python3 --version

0.2 Create + activate venv

python3 -m venv .venv
source .venv/bin/activate

Windows: .venv\Scripts\activate

🔐 Install + License

0.3 Install Rasa Pro

pip install rasa-pro

0.4 Request free educational/dev license

rasa.com/rasa-pro-developer-edition-license-key-request

0.5 Export license

export RASA_LICENSE="<the license string emailed to you>"

RASA_PRO_LICENSE still works (deprecated).

✅ Verify Installation

rasa --version

Expect: - Rasa Pro version - SDK version - Python version - License Expires date

If error: license.not_found.error → license variable not set in current shell.

📦 Part 1 — Pull Starter Project

rasa init --template tutorial

Why tutorial template? - Uses free hosted Rasa tutorial LLM - No OpenAI billing/API-key setup needed for students

🗂️ What the Scaffold Includes

File/Folder	Purpose
config.yml	CALM config (CompactLLMCommandGenerator, FlowPolicy)
domain.yml	Slots + responses
data/flows.yml	Conversation business logic
data/patterns.yml	Built-in repair flows
actions/	Custom Python actions
credentials.yml	Channel connectors (REST default)
endpoints.yml	LLM/action server endpoints

Frame this scaffold as a map of responsibilities, not a collection of files students need to memorize. For example, if a student wants the FAQ assistant to answer “Where is the lab?”, they would add the conversation logic in data/flows.yml and the response text in domain.yml; config.yml controls how CALM interprets requests, while credentials.yml and endpoints.yml connect the assistant to channels and services. Ask students which file they would inspect first when changing an answer, then trace the request from user message to flow to response.

🧩 Part 2 — Core CALM Concepts

• description: flow trigger intent (LLM-readable)
• collect: request and fill slot
• Slots typed memory (text, categorical, float, …)
• Patterns built-in repair/handoff/chitchat flows

⭐ Key feature for demo: digressions (pause current flow, handle interruption, then resume)

Explain that the description is a natural-language guide for recognizing when a flow is relevant, while the flow’s steps still control what the assistant does. For example, “I need help with my laptop” can trigger an IT-help flow; collect asks for and stores a needed detail such as the issue category. Then demonstrate a digression: while describing a software problem, the user asks “Actually, what are your office hours?” The assistant can answer that side question and return to the unfinished help flow. Emphasize that this is useful only when returning to the original task feels clear and expected to the user.

🟢 Part 3 — Scenario 1: Course FAQ Assistant

Goal: Trigger flows + static responses (no slot collection)

Add flows in data/flows.yml

flows:
  tell_office_hours:
    description: User wants to know instructor office hours.
    steps:
      - action: utter_office_hours

  tell_syllabus_location:
    description: User wants to know where to find the course syllabus.
    steps:
      - action: utter_syllabus_location

  tell_assignment_deadline:
    description: User wants to know when an assignment is due.
    steps:
      - action: utter_assignment_deadline

🟢 Scenario 1 — Responses in domain.yml

responses:
  utter_office_hours:
    - text: "Office hours are Tue/Thu 2–3pm in Ford 210, or by appointment."
  utter_syllabus_location:
    - text: "The syllabus is posted on the course Canvas page, under Files."
  utter_assignment_deadline:
    - text: "Check the Canvas calendar for exact due dates — they vary by assignment."

⚠️ Keep one responses: block only.

🧪 Scenario 1 — Validate, Train, Test

rasa data validate
rasa train
rasa shell

Visualization:

rasa inspect

Try paraphrased user asks to demonstrate CALM generalization.

🚀 Scenario 1 — Deploy Locally

rasa run --enable-api --cors "*"

Health check:

curl -s http://localhost:5005/status

REST test:

curl -s -X POST http://localhost:5005/webhooks/rest/webhook \
  -H "Content-Type: application/json" \
  -d '{"sender": "test_user", "message": "when are office hours?"}'

💬 Browser Demo

open "webchat_test.html"

• Real chat UI over REST endpoint
• Refresh creates new random sender/session
• Great for showing “deployed product” feel in class

🟡 Part 4 — Scenario 2: IT Help-Desk Assistant

Goal: slot collection + conditional branching + digressions/corrections

Add slots (domain.yml)

slots:
  issue_category:
    type: categorical
    values: ["hardware", "software", "account_access", "other"]
    mappings:
      - type: from_llm
  issue_description:
    type: text
    mappings:
      - type: from_llm
  urgency:
    type: categorical
    values: ["low", "medium", "high"]
    mappings:
      - type: from_llm

🟡 Scenario 2 — Add Responses

responses:
  utter_ask_issue_category:
    - text: "What kind of issue is this — hardware, software, or account access?"
  utter_ask_issue_description:
    - text: "Briefly describe the problem."
  utter_ask_urgency:
    - text: "How urgent is this — low, medium, or high?"
  utter_escalate:
    - text: "This looks urgent — flagging it for immediate review."
  utter_ticket_summary:
    - text: "Ticket created: {issue_category} issue — '{issue_description}', urgency: {urgency}."

🟡 Scenario 2 — Flow with Branching

submit_it_ticket:
  description: User wants to report a tech problem or request IT help.
  steps:
    - collect: issue_category
    - collect: issue_description
    - collect: urgency
      next:
        - if: slots.urgency == "high"
          then:
            - action: utter_escalate
            - link: pattern_human_handoff
        - else:
            - action: utter_ticket_summary
              next: END

🔍 Scenario 2 — What to Test

1. Clean run: fill all slots → ticket summary
2. Digression: interrupt mid-flow, ask ticket status, resume prior flow
3. Correction: change urgency after initial value

Commands:

rasa data validate
rasa train
rasa inspect

🧯 Part 5 — Troubleshooting Cheat Sheet

Symptom	Cause	Fix
duplicate flows/responses key	second top-level header added	keep one block
CategoricalSlot ... unexpected keyword	slot indentation error	align slot names
missing a 'next' in branch	terminal step without next	add next: END
invalid choice: 'validate'	wrong command	rasa data validate
address already in use	port conflict	free port/process
license.not_found.error	env var unset	export license

⚙️ Part 6 — Optional: Switch LLM Provider

Default tutorial endpoint is free.
Swap in OpenAI or Gemini in endpoints.yml when needed.

provider: openai
model: gpt-5.1-2025-11-13
api_key: ${OPENAI_API_KEY}

Cost note: paid providers bill every class run.

✅ Deliverables for Today

• Working Scenario 1 assistant (FAQ)
• Working Scenario 2 assistant (ticket flow + branching)
• Successful local deployment test (REST + browser)
• Error-fix evidence from validate/train cycle

Next

• Dockerize for persistent deployment
• Later: visualize/edit same artifacts in Rasa Studio