Scenario
A support bot uses vector search to find the knowledge-base article most relevant to a question. Users complain it keeps surfacing an unrelated article as the top result, even when a clearly better article exists.
cosine_distance returns a distance, where 0 means identical and larger means less similar. search embeds the query, scores every document, and returns the top results. The embeddings here are tiny hand-made vectors so you can reason about them by eye.
Run the code as-is. The query is about refunds, and the first result is the shipping article.
Your task: find why the ranking is backwards and fix search. Do not change the documents, the embeddings, or cosine_distance.
Interviewers love this one because the code runs without errors and returns plausible-looking output. The only way to catch it is to reason about what the numbers mean.