The RAG architecture is divided into a retrieval stage and a generation stage; the retrieval stage uses vector databases (such as Chroma, Pinecone, and Milvus) to recall relevant text segments via cosine similarity.