# Quiz: Connect company information, live records and testing

## Student quiz

### 1. An employee asks how many vacation days can be carried forward. The policy uses the words “annual leave.” What helps a RAG system find the relevant passage despite the wording difference?

A. Comparing numerical representations of meaning so related passages can be retrieved
B. Searching only for the exact words used in the question
C. Giving the model permission to change the policy
D. Asking the employee to rewrite the question with the policy's exact words

**Correct answer: A.** In the taught example, embeddings represent meaning and help retrieve a related passage. B depends on an exact wording match and can miss related meaning. C is unnecessary and unsafe for a question-answering job. D changes the employee's wording instead of letting the system retrieve a related passage.

### 2. An assistant receives a contract as context for one answer. Which concern remains even if the provider says it does not use the file for training?

A. Whether the model can generate tokens
B. Whether an API exists anywhere on the internet
C. Whether every future answer must quote the contract
D. Where the file and conversation are stored, who can access them and when they are deleted

**Correct answer: D.** A no-training statement addresses only training use. Storage, access, retention and other recipients remain separate questions. A and B do not answer the information-handling concern. C invents a requirement unrelated to retention.

### 3. Priya wants a repeatable first test for an HR assistant. Which plan gives her the most useful evidence?

A. Run only easy questions chosen by the seller and judge whether the answers sound professional.
B. Save checkable questions, sources and expected results, include missing or conflicting information, and keep failed cases for reruns.
C. Ask one difficult question repeatedly until the assistant gives the preferred wording.
D. Test with real sensitive records and allow every available action.

**Correct answer: B.** A repeatable evaluation needs prepared evidence, expected outcomes and difficult cases that remain available after a change. A tests polish rather than correctness. C can reward a preferred sentence without locating the requirement. D creates unnecessary exposure and ignores the fictional-record boundary taught for tool tests.