Advanced RAG interview questions

Basic RAG does one similarity search on the user's raw question and hopes the top-k chunks are right. In practice the retrieval is what breaks: a question is worded nothing like the document that answers it, key facts get split across separate chunks, and a single search only ever asks once. A great model handed the wrong passages still gives a wrong answer, so advanced RAG focuses almost entirely on getting the right text in front of the model.

All three fix the query before searching. Rewriting rephrases a vague or messy question into something cleaner. Multi-query turns one question into several rewordings, searches with each, and unions the results, so a single bad phrasing doesn't sink you. HyDE (Hypothetical Document Embeddings) goes further: the model drafts a fake answer to the question and searches with that, because an answer looks far more like the real document than a question does.

Vector search returns whatever is nearest in meaning-space, and a question ("why is onboarding slow?") lands far from the paragraph that answers it. A drafted answer uses the same vocabulary and shape as the real document, so it sits right next to it — the search then retrieves the true passages. The draft doesn't have to be correct; it only has to be shaped like the target. The risk is that on a niche topic the model can invent a wrong premise and steer the search off.