Recall = (relevant retrieved) / (total relevant in catalog). High recall means you’re finding most of the relevant items; low recall means you’re missing some. For ecommerce, low recall is invisible to shoppers (they don’t know what they didn’t see) but shows up as missed revenue, especially on long-tail queries.
Recall@K is the operational version, often computed against a held-out set of (query, known-relevant-product) pairs from purchase data. R@100 = 0.95 means 95% of the time, the relevant product is somewhere in the top 100 — a healthy floor for a hybrid retriever before re-ranking.
In two-stage search, the retriever optimizes recall (don’t miss relevant items in the top-N candidate pool); the re-ranker optimizes precision (put the best candidates first). Conflating the two leads to bad architecture decisions.