GenAI Field Guide

Agents and beyond

RAG, retrieval-augmented generation

Instead of answering from what it memorized in training, the model is handed the relevant chunks of your documents at question time and told "answer from these."

≈2 minLandscape facts checked — re-verify before relying on them.
1 · Chunk Split yourdocuments intopassages 2 · Embed Turn each chunkinto a meaningvector 3 · Retrieve Find the chunksnearest thequestion 4 · Generate Answer from thosechunks, withcitations
Chunk, embed, retrieve, generate.

How it works, in four steps

  1. Chunk. Split your documents into pieces.
  2. Embed. Turn each chunk into an embedding, a list of numbers that captures its meaning.
  3. Retrieve. At question time, find the chunks nearest to the question.
  4. Generate. Put those chunks in the prompt with "answer from these," and get citations you can check.

When to use what

Situation Use Why
Document fits in context and is used once Just attach it Simplest; the model sees everything
Corpus is large, changes often, or many users query it RAG Only relevant chunks are sent; stays current without retraining; citations
You need the model to speak differently (tone, format), not know something Fine-tuning Advanced, rarely the right first tool

Novice on-ramp: NotebookLM (consumer Gemini) is RAG with a friendly face. On GenAI.mil, a Project with files is RAG-lite.