Building AI Chatbots with Rasa

1. Architecture Overview (Rasa Stack)

Rasa consists of two primary subsystems:

• Rasa NLU → intent classification + entity extraction
• Rasa Core → dialogue management (policies + state machine)

Key internal components:

• Pipeline (NLU processing chain)
• Stories / Rules (dialogue supervision)
• Domain (schema: intents, entities, slots, actions)
• Tracker (conversation state)
• Policies (decision logic)

2. Environment Setup

2.1 System Requirements

• Python 3.8–3.11 (strict compatibility matters)
• pip / venv (or Conda)
• Optional: Docker for containerization

2.2 Create Virtual Environment

python -m venv rasa-env
source rasa-env/bin/activate   # Linux/macOS
rasa-env\Scripts\activate      # Windows

2.3 Install Rasa

pip install rasa

Verify:

rasa --version

3. Initialize a Rasa Project

rasa init

This generates:

.
├── data/
│   ├── nlu.yml
│   ├── stories.yml
│   └── rules.yml
├── domain.yml
├── config.yml
├── actions/
│   └── actions.py
├── endpoints.yml
└── credentials.yml

4. NLU Pipeline Configuration

Edit config.yml.

Example pipeline (DIET-based modern setup):

pipeline:
  - name: WhitespaceTokenizer
  - name: RegexFeaturizer
  - name: LexicalSyntacticFeaturizer
  - name: CountVectorsFeaturizer
  - name: DIETClassifier
    epochs: 100
  - name: EntitySynonymMapper

Explanation:

• Tokenizer → splits input text
• Featurizers → convert text → vectors
• DIETClassifier → multitask transformer (intent + entities)

5. Define Domain

domain.yml is the schema layer.

intents:
  - greet
  - goodbye
  - ask_weather

entities:
  - location

slots:
  location:
    type: text

responses:
  utter_greet:
    - text: "Hello! How can I help?"

actions:
  - action_get_weather

6. Create Training Data

6.1 NLU Data (data/nlu.yml)

nlu:
- intent: greet
  examples: |
    - hi
    - hello
    - hey there

- intent: ask_weather
  examples: |
    - what's the weather in Zagreb
    - weather in London

6.2 Stories (data/stories.yml)

stories:
- story: weather path
  steps:
  - intent: ask_weather
  - action: action_get_weather

6.3 Rules (data/rules.yml)

rules:
- rule: respond to greeting
  steps:
  - intent: greet
  - action: utter_greet

7. Custom Actions (Business Logic)

Install SDK:

pip install rasa-sdk

Edit actions/actions.py:

from rasa_sdk import Action
from rasa_sdk.events import SlotSet

class ActionGetWeather(Action):
    def name(self):
        return "action_get_weather"

    def run(self, dispatcher, tracker, domain):
        location = tracker.get_slot("location")
        dispatcher.utter_message(text=f"Weather in {location} is sunny.")
        return []

Run action server:

rasa run actions

8. Training the Model

rasa train

Artifacts:

models/
└── model.tar.gz

9. Testing the Chatbot

9.1 Interactive Shell

rasa shell

9.2 Test NLU Only

rasa shell nlu

9.3 Automated Testing

rasa test

10. Dialogue Policies

In config.yml:

policies:
  - name: RulePolicy
  - name: MemoizationPolicy
  - name: TEDPolicy
    max_history: 5
    epochs: 100

Key notes:

• TEDPolicy → transformer-based dialogue prediction
• MemoizationPolicy → exact story recall
• RulePolicy → deterministic flows

11. Slots and Context Handling

Slots maintain conversation memory.

Example:

slots:
  location:
    type: text
    influence_conversation: true

Used by policies for context-aware predictions.

12. Forms (Structured Conversations)

Example use case: collecting user info.

forms:
  weather_form:
    required_slots:
      - location

13. Integrations (Channels)

Configure credentials.yml:

rest:

14. Deployment Options

14.1 Local Server

rasa run --enable-api

14.2 Docker Deployment

docker run -p 5005:5005 rasa/rasa:latest

15. Tracker Store & Persistence

In endpoints.yml:

tracker_store:
  type: SQL
  dialect: "postgresql"
  url: "localhost"
  db: "rasa"

17. Evaluation & Metrics

rasa test nlu
rasa test core

Outputs:

• Precision / Recall / F1
• Confusion matrix
• Story accuracy

21. Minimal Production Blueprint

User → Channel (Slack/Web)
     → Rasa Server (NLU + Core)
     → Action Server (Python logic)
     → External APIs (weather, DB, etc.)

22. When to Use Rasa

Use Rasa if:

• You need on-premise NLP
• You require custom dialogue control
• Data privacy is critical

Avoid if:

• You want plug-and-play SaaS (consider alternatives like Dialogflow)