AIF-C01 Practice Question 66: Retrieval Relevance Reranking
AWS Certified AI Practitioner (AIF-C01) · Test 7 · question 6 of 10
A RAG assistant over a 200,000-document policy library keeps answering from passages that merely share vocabulary with the question instead of actually answering it. Retrieval already returns 20 chunks per query, and the generation model sometimes fixes on the wrong one. Which change most directly improves which passages reach the model?
Show answer and explanation
Answer:
A reranker model *calculates the relevance of chunks to a query and reorders the results based on the scores that it calculates*. Reranker models are *trained to identify relevance signals based on a query and then use those signals to rank documents*, which is a sharper judgement than the vector similarity that produced the original 20 hits. AWS notes you can therefore *retrieve fewer, but more relevant, results*, which also decreases cost and latency, and that reranking *overrides the default ranking that Amazon Bedrock Knowledge Bases determines* (option D).
- A is wrong because temperature governs randomness when sampling the next token. It changes how the answer is worded, not which passages were retrieved or how the model weighs them. - B is wrong because the problem is precision, not recall; the right passage is probably already in the 20. Pushing 100 passages into the context adds noise, cost and latency, and makes "fixes on the wrong one" more likely rather than less. - C is wrong because chunk size changes how the corpus is cut up, not how relevance is scored. Smaller chunks can even make things worse by splitting a complete answer across several passages.
Reference: Improve the relevance of query responses with a reranker model in Amazon Bedrock