A candidate pool is the intermediate set of items retained as potential recommendations before expensive scoring and final ranking. Production systems build it because scoring millions of catalog items for every user is impractical. section 1.4
Retrieval methods progressively narrow the catalog into a pool of hundreds or thousands. Later stages score, re-rank, and filter these items, so the pool must be broad enough to contain relevant items without creating unnecessary downstream cost. section 1.4 section 4.6.5
Evaluate and operate the pool by separating retrieval concerns from ranking concerns:
- Measure retrieval recall: determine whether relevant items entered the pool, rather than judging their internal order. section 4.6.5
- Test several pool sizes, such as 50, 100, 200, and 500, to locate the recall elbow and choose a cap. section 4.6.5
- Treat filtering as correctness: remove consumed items, unavailable items, and items that violate business rules or content policies. section 4.6.5
A retrieval pool
A retrieval stage returns 500 items from a catalog of potentially millions. If all 10 items a user would engage with appear anywhere in that pool, retrieval has succeeded even when the remaining items are mediocre; the scoring stage can reorder them afterward. section 4.6.5
