RAG Fundamentals — Final Exam

What core problem does RAG primarily address?

You need the model to answer in your company's brand voice, but it already knows the facts. Which tool fits best?

What is an embedding?

Which similarity metric is most commonly used to compare embeddings in RAG retrieval?

Why must queries and documents be embedded with the same model?

Which of the following are risks of making chunks TOO LARGE? (Select all that apply)

Which stages happen at QUERY time (per request) rather than during offline indexing? (Select all that apply)

What is the purpose of overlap when chunking documents?

In the generation step, what single instruction most distinguishes grounded RAG from just pasting search results next to an LLM?

Which metric tells you whether retrieval fetched the chunks that actually contain the answer?

The retrieved context clearly contains the correct answer, but the model's answer is invented and unsupported. Which metric captures this failure?

A RAG answer is wrong. What is the most useful FIRST diagnostic step?

Name the two-word technique (acronym expanded or not) for grounding an LLM in retrieved external documents at query time.

In one word, what failure mode does instructing the model to 'say you don't know' primarily prevent?