AI 101

RAG

Also known as: retrieval-augmented generation

Inference

RAG (Retrieval-Augmented Generation) is a pattern: before answering, the AI retrieves — searches — the documents you care about, stuffs the relevant parts into the prompt, and then answers.

The pipeline in three moves:

  1. Chunk & index — your documents (notion, wiki, PDFs, code) are cut into pieces and indexed so they can be searched quickly.
  2. Retrieve — when you ask a question, the system finds the ~5 most relevant chunks.
  3. Generate — the model answers using those chunks as notes: “Here are the relevant excerpts — answer the question based on them.”

Why it exists:

  • A model only knows what was in its training data — nothing about your company, your server, your 2025 contract.
  • Its context window can’t hold your entire wiki.
  • RAG bridges both gaps: the model gets fresh, private, specific material — and can even cite which page it used.

RAG vs fine-tuning, in one line: RAG gives the model *notes to consult* (instant, cheap, updatable); fine-tuning changes the model's *brain* (slow, permanent). Most "my AI knows my company" use cases are RAG, not fine-tuning.