Make Recall a Product Requirement
Define retrieval quality using ranked-result metrics such as precision@k, recall@k, MAP, or NDCG.
The move: score the ranked list before you praise the generated answer. A vector database is usually one component in a retrieval pipeline. The database returns candidates. A reranker may reorder them. A generator may write an answer. If you only judge the final text, you cannot tell which layer failed. Precision At K Use when the first screen or first few citations must be clean. High precision means fewer distractors in the slots users or models actually consume. Recall At K Use when missing the right source is costly. High recall means the relevant record appears somewhere in the…
Sign up free — one personalized lesson every day, matched to your role and goals.
Already have an account? Sign in