How akshatt/receptionbot works
Project Type and Purpose
• ReceptionBot is a conversational AI agent built on the Rasa framework (NLU + Core).
• It handles hotel-related intents: booking rooms, scheduling cleaning, FAQs, dynamic slot filling, and interruptions mid-flow.
Top-Level Components
A. Offline Training Pipeline
B. Runtime Serving Architecture
Offline Training Pipeline
• Training Data Store (folder “data/” + domain.yml + responses.md)
• Configuration (config.yml)
• Rasa Trainer
– NLU pipeline (DIETClassifier, tokenizers, featurizers)
– Core policies (Memoization, TED, RulePolicy)
• Output (models/ timestamped .tar.gz)
data/ + domain.yml + config.yml → Rasa Trainer → models/
Runtime Serving Architecture
1. Channel Connectors (from credentials.yml & endpoints.yml)
– e.g. CLI (rasa shell), REST input, Facebook Messenger, etc.
2. Rasa Server
• Rasa NLU component (intent/entity recognition)
• Rasa Core component (dialogue management using policies + tracker store)
• Domain definitions (responses, slots, forms, actions)
3. Action Server (actions.py)
– Custom action endpoints (external API calls, business logic)
4. Tracker Store (optional persistent storage)
– Could be InMemory, Redis, or a database (configured in endpoints.yml)
5. Response Output
– Bot replies back to the channel connector which forwards to the user UI
User → Channel Connector → Rasa Server → (NLU→Core) →
– If custom action required → Action Server → Core
– Core → Response → Channel Connector → User
Component Details & Technologies
• Rasa Framework v1.8 (Python 3.6.8)
• config.yml defines pipeline components (e.g., DIETClassifier, CountVectorsFeaturizer) and Core policies
• domain.yml declares intents, entities, slots, forms, responses, action names
• data/nlu.md, data/stories.md, data/responses.md encapsulate training examples
• endpoints.yml configures action endpoint URL and tracker store
• credentials.yml lists channel credentials (shell, REST, messenger, etc.)
• actions.py hosts Python functions for custom actions (e.g., scheduling cleaning, booking logic)
Architectural Patterns & Principles
• Separation of concerns: NLU vs. Core vs. custom actions
• Modular data configuration: domain, stories, NLU examples, responses in distinct files
• Pluggable pipelines: config.yml allows swapping models or policies
• Stateless core with optional persistent tracker store for conversational context
Capturing Details
• In the Training section, annotate that multiple timestamped models appear under models/
• In the Runtime section, note that “rasa run actions” launches the Action Server separately
• Split between Rasa NLU processing and Rasa Core dialogue management
• How interruptions mid-flow are handled by the same Core policies and fallback rules
[OFFLINE TRAINING – Blue Section]
data/nlu.md, data/stories.md, domain.yml, config.yml
↓
Rasa Trainer
↓
models/
[RUNTIME SERVING – Green Section]
User ←→ Channel Connector(s) ←→ Rasa Server
↳ Domain + policies (domain.yml, config.yml)
↳ ←→ Tracker Store
↳ → Custom Actions (Action Server via endpoints.yml)
↓
Bot Responses