Conversational AI Design
Design conversational AI systems with Rasa 3.x NLU pipelines, dialogue management, and LLM-based chatbot patterns.
Rasa NLU Pipeline
Configuration
language: en
pipeline:
- name: WhitespaceTokenizer
- name: RegexFeaturizer
- name: LexicalSyntacticFeaturizer
- name: CountVectorsFeaturizer
- name: DIETClassifier
epochs: 100
- name: EntitySynonymMapper
- name: ResponseSelector
epochs: 100
Intent & Entity Design
- Intents: User goals (e.g.,
greet,book_flight,check_status) - Entities: Data to extract (e.g.,
date,location,order_id) - Minimum 10 examples per intent for reliable classification
Dialogue Management
Policy Stack
policies:
- name: RulePolicy # Handle explicit rules
- name: TEDPolicy # ML-based dialogue
epochs: 100
- name: MemoizationPolicy # Exact conversation matches
max_history: 5
Conversation Patterns
- Form-based: Collect structured data (bookings, orders)
- FAQ-style: Direct question → answer
- Multi-turn: Context-aware follow-ups
- Fallback: Handoff to human when confidence < threshold
LLM Chatbot Patterns (LangChain)
Conversational Memory
from langchain.memory import ConversationBufferMemory
memory = ConversationBufferMemory(
memory_key="chat_history",
return_messages=True,
max_token_limit=2000
)
RAG for Chatbots
from langchain.chains import RetrievalQA
qa_chain = RetrievalQA.from_chain_type(
llm=llm,
retriever=vector_store.as_retriever(),
chain_type="stuff",
memory=memory
)
Best Practices
- Always confirm before destructive actions
- Provide options not open-ended questions when possible
- Handle fallbacks gracefully ("I didn't understand. Try: X, Y, Z")
- Log conversations for analysis and improvement
- Test with real users — not just developers
- Set expectations — tell users what the bot can/can't do
Anti-Patterns
- ❌ No fallback handling
- ❌ Overly long responses
- ❌ Pretending to be human
- ❌ Ignoring context in multi-turn
- ❌ No escape hatch to human agent