The core idea

Fetch the best chunks, then build the prompt.

Retrieval is the live step: embed the question, pull the top matching chunks, and assemble them into the context the model answers from.