A working RAG system runs four stages. Chunking splits documents into passages small enough to search. Embedding turns each passage into a vector, a list of numbers that represents its meaning. Retrieval embeds the question the same way and pulls back the passages whose vectors are closest. Generation places those passages into the prompt and produces an answer grounded in them. The first two stages happen ahead of time; the last two happen per question.
How AI Answers Questions About Documents It Never Learned