RAG retrieves relevant documents from a vector database using semantic similarity search injects them into the LLM context and generates a response grounded in the retrieved content. Main failure modes are retrieval failures context window overflow and hallucinations about retrieved content.