Exam-style
An evaluation run shows the correct passage sits inside the top 25 retrieved chunks for 90 percent of questions, but inside the top 3 passed to the model for only 40 percent. What change lifts answer quality most directly?
Reveal the answer
B. Retrieve 25 candidates and rerank them with the Amazon Bedrock Rerank API, passing the top 3 reranked chunks on
Recall is already good, since the right passage is in the top 25 nine times out of ten. The ordering is what fails, and reranking reorders. The Rerank API takes the query and up to 1,000 candidate documents in one request, scores each candidate against the query text rather than comparing two precomputed vectors, and returns them in relevance order. Passing all 25 chunks instead adds input tokens to every call and buries the right passage among 22 irrelevant ones. The other choices address recall or chunk shape, and neither is failing here.
Q. The right chunk is retrieved but ranks twelfth, below the cutoff. What promotes it?
A. A reranker. The Amazon Bedrock Rerank API re-scores the top-N candidates against the query text, more accurately than the first-stage embedding score. It supports two models, Amazon Rerank 1.0 (amazon.rerank-v1:0) and Cohere Rerank 3.5 (cohere.rerank-v3-5:0), and in us-east-1 only the Cohere one is available. Retrieve wide, rerank narrow.
Why? Reranking reorders what retrieval already found. If the chunk is not in the top-N at all, fix retrieval first.