Describe a basic RAG architecture and its two main components

WHAT THIS TESTS: Whether the candidate can decompose a RAG system into its two fundamental subsystems and explain the data flow and control flow between them. Interviewers use this to verify you understand that RAG is an architecture pattern rather than a specific model, and that you can reason clearly about where information lives at inference time.

A GOOD ANSWER COVERS: First, name the two components explicitly: the retriever and the generator. Second, describe the retriever as the module that accepts a user query, searches an external knowledge base such as a vector database or document store, and returns the top-k relevant chunks or passages. Third, describe the generator as the large language model that receives both the original query and the retrieved context, then synthesizes a coherent, factually grounded response using that evidence. Fourth, mention the orchestration step that assembles the retrieved results into the model context window, because senior interviewers want to see you understand the interface between the parts.

COMMON WRONG ANSWERS: Confusing RAG with fine-tuning by claiming the external documents are baked into the model weights. Describing only the LLM and forgetting the retriever entirely. Saying the retriever is just a web search without mentioning an indexed document corpus. Overcomplicating the answer with embedding model details before establishing the high-level split.

LIKELY FOLLOW-UPS: How do you measure retrieval accuracy? What happens if the retriever returns irrelevant chunks? How do you handle context window limits when feeding documents to the generator? When would you choose RAG over fine-tuning for domain adaptation?

ONE CONCRETE EXAMPLE: Imagine an internal company chatbot built for HR questions. The retriever queries a vector index of Confluence pages and Slack archives to find five relevant snippets about the vacation policy. The generator, a hosted LLM, receives the question plus those five snippets and drafts an answer citing the specific sources. If the HR policy changes tomorrow, you update the index, not the model weights.