RAG Architecture Knowledge Quiz

1.Which chunking strategy preserves the most semantic coherence for technical documentation?

Fixed-size 512-token windows with no overlap.Recursive splitting along structural markers (headings, lists, code blocks).One chunk per document.Splitting by sentence boundaries only.

2.When does a sparse retriever (BM25) tend to outperform a dense embedding retriever?

When queries involve rare exact tokens like product SKUs or error codes.For long open-ended natural-language queries.When the corpus is small (< 1k documents).Never — dense retrievers always win.

3.What is the primary role of a cross-encoder reranker in a RAG pipeline?

Reduce embedding storage costs.Re-score the top-k retrieved candidates jointly with the query for relevance.Generate the final answer for the user.Replace the embedding retriever entirely.

4.Which evaluation metric is most useful for diagnosing retrieval-vs-generation failure modes?

End-to-end answer correctness only.Retrieval recall@k separately from answer faithfulness.User-perceived response time.Average response length.

5.Why doesn't 'just use a longer context window' eliminate the need for retrieval?

Long-context models always cost less per token.Models still degrade on relevance ranking inside long contexts, and latency + cost scale poorly.Context windows can't fit more than 32K tokens in practice.Retrieval is required for compliance reasons.