## TL;DR
Conversational AI spans from task-oriented bots that book flights to open-domain chatbots that discuss any topic. Modern LLMs have collapsed the traditional modular pipeline into end-to-end neural approaches, while frameworks like Rasa CALM bring enterprise-grade reliability.

## Core Explanation
Traditional task-oriented pipeline: NLU (intent classification + entity extraction), DST (dialogue state tracking — what has been said and confirmed), DP (dialogue policy — what to do next), NLG (generate response). Example: "Book a flight to Paris on Friday" → intent=book_flight, destination=Paris, date=Friday. LLM-based approaches use function calling to interact with APIs and databases within conversational flow.

## Detailed Analysis
Dialogue state tracking tracks slot values (confirmed/requested), dialogue acts (inform, request, confirm, greet). Modern LLM approaches use tool-augmented generation: the LLM decides when to call search, database, or booking APIs. Rasa CALM maps business logic to natural language flows with guardrails. Chatbot evaluation: task completion rate, user satisfaction, conversation length, and hallucination rate.

## Further Reading
- Rasa Open Source Documentation
- Google Dialogue Flow CX
- ParlAI (Facebook AI Research)