What retrieval does
Retrieval turns a request into a few hundred candidate items pulled from the entire catalog in milliseconds. It cannot afford a heavy per item model, so it leans on precomputed structures and cheap lookups.
Common retrieval sources
- Co occurrence lists, like users who watched this also watched that.
- Embedding nearest neighbors, where a user vector finds nearby item vectors.
- Popularity and trending items as a strong fallback.
- Rule based sources such as recent searches or followed creators.
Blending many sources
Real systems run several retrieval sources in parallel and union their results. Each source covers a different intent, so blending raises overall recall. Duplicates are merged before ranking sees the pool.
The recall target
Retrieval is judged by recall at k: of the items the user would actually engage with, how many appear in the candidate set. A high recall pool gives ranking room to shine; a thin pool caps the whole system.
Key idea
Candidate retrieval blends several cheap sources to build a high recall shortlist quickly, because ranking can only reorder what retrieval supplies.