How Conversational AI Works: Intents, NLU & LLMs
Every AI companion, voice assistant and chatbot rests on the same foundation: conversational AI. This article explains how it works — from the classic pipeline of intents and entities that frameworks like Violet were built around, to the large-language-model approach that powers today's AI girlfriend apps.
The classic pipeline: NLU, intents, entities
Before LLMs, conversational systems worked as a pipeline. A user's message was passed through Natural Language Understanding (NLU), which produced two things:
- Intent — the goal behind the message (e.g.
getWeather,playMusic). - Entities — the parameters (e.g. a city, a date, a song title).
A dialogue manager then decided what to do: fill missing slots by asking follow-up questions, call your application logic, and produce a response. This is exactly the model Violet made easy to script — you defined what the app expected and how to resolve each turn.
Dialogue management and state
The hard part of a real conversation is state: remembering what has already been said and what the system is still waiting for. Classic engines tracked this with slots and goals. Violet, for instance, saved each user's conversation state in a session so multi-turn goals worked out of the box.
The LLM shift
Large language models changed the picture. Instead of classifying each message into a fixed intent, an LLM generates the next reply directly from the conversation so far plus a system prompt that defines the persona. This makes open-ended, natural conversation possible — the leap that made AI companions convincing.
Modern systems are usually hybrids: the LLM handles free-form chat, while intent-style routing still triggers concrete actions like generating an image or updating a setting. If you want to build one yourself, follow our tutorial on building an AI chatbot with Node.js.
Why memory is the real differentiator
An LLM on its own is stateless. What makes a companion feel like a relationship is a memory store — long-term facts (your name, preferences, past conversations) retrieved and fed back into the model on every turn. The quality of that memory is the single biggest factor separating a forgettable bot from one people pay for. We break this down in how AI girlfriend apps work.
Frequently asked questions
What is an intent?
The goal behind a message — 'book a flight', 'tell me a joke'. Classic systems classify into fixed intents; LLMs infer intent flexibly from context.
Do AI companions still use intents?
Partly — they're LLM-first for conversation but still use intent-like routing for actions such as image generation.
What makes a conversation feel continuous?
Memory: a store of long-term facts fed back to the model each turn.
Next: Build an AI Chatbot with Node.js → · Back to the AI Companions Guide