Design a retrieve-and-rerank pipeline that balances recall, precision, latency, and context budget.
The right passage often appears in top 30 vector candidates, but not in the top 6 chunks sent to the generator. First stage retrieves for recall; second stage reranks for precision in the final slots. The common trap is adding a reranker without checking candidate recall. A reranker cannot promote a passage that the vector database never returned. Measure candidate recall Run the current retriever at candidate_k=40 and check whether relevant records appear in the candidate set. This proves whether the first stage gives the reranker a chance to succeed. Rerank candidates Score query-candidate pairs with a reranker and reorder…
Sign up free — one personalized lesson every day, matched to your role and goals.
Already have an account? Sign in