RAG gets explained like an architecture. Diagrams, boxes, arrows, and a vocabulary to go with them.
Here is the whole thing. The model does not know your documents. So before you ask, you find the relevant bits and paste them in above the question. That is it.
Everything else is a tweak on those two moves. Chunking is how big the pasted bits should be. Embeddings are how you find them without exact keywords. Reranking is picking three out of ten. Real problems, but none of them change the shape.
This matters when something breaks. Your first question should be: did the right text actually make it into the prompt? Usually it did not, and no amount of prompt tuning downstream will fix that.
I have watched people spend a day on temperature and system prompts when their search was returning the wrong document every time. The diagram had convinced them the problem was somewhere clever.
So most of the work is search quality. That sounds like a letdown. It is not. Search is a well-understood field, and you get to use all of it.
Good search, good RAG. Bad search, and you have built an expensive way to be confidently wrong.