Why RAG exists, the full pipeline from chunking to grounded generation, embeddings as the retrieval mechanism underneath it, chunking trade-offs, evaluating retrieval quality, and advanced patterns like hybrid search and re-ranking.
A language model's parametric memory is fixed at training time and cannot be traced back to a source — RAG solves this by retrieving relevant passages from an external knowledge source at query time and injecting them directly into the prompt, so answers are grounded in text the model can actually see. This module builds the full pipeline end to end: chunking documents, generating embeddings, storing and searching vectors, retrieving the right passages, and generating an answer conditioned on that retrieved context, plus the tuning decisions (chunk size and overlap, top-k) that determine whether the pipeline actually works.
By the end of this module you should be able to explain why RAG reduces hallucination risk without eliminating it, why retrieval precision and recall have to be measured separately from final-answer quality, and how retrieval has to fit inside the same context-window budget Module 5 covers — the exact foundation Project 1's Customer Support RAG Chatbot, later in this program, is built on.