RAG (Retrieval-Augmented Generation).

Retrieval-Augmented Generation (RAG) is how production AI applications answer questions about specific business data. The pattern: when a user asks a question, the system first retrieves relevant information from your knowledge sources (databases, documents, systems), then gives both the question and the retrieved data to the LLM. The LLM answers based on the retrieved data, not what it happens to remember from training.

Why this matters: out-of-the-box LLMs know what they were trained on — generally the public internet up to some cutoff date. They don't know your jobs, your customers, your past quotes, your inventory levels. RAG bridges that gap without the cost and complexity of training a custom model. You keep the LLM, you add a retrieval system over your own data, and suddenly the AI can answer "what did we quote Acme in March?" or "where's job 1647 right now?".

For practical SMB applications, RAG is the foundation of internal AI assistants. The assistant doesn't memorise your business — it queries your business when asked. That has three important properties: answers reflect current state (not a frozen snapshot), data stays in your environment (the model only sees retrieved snippets to answer one question), and you can update the underlying data freely (no retraining needed).

Deployment cost is much lower than fine-tuning, and the build is much faster. Most useful internal AI assistants are RAG over a structured database and a document store — usually 4-8 weeks of build for the first version.