Recommendation Systems

A recommendation system suggests items — products, articles, videos, songs — a user is likely to want next. Classical approaches split into collaborative filtering (“users like you also liked…”) and content-based filtering (“this is similar to what you just viewed”). Embeddings modernize both: represent every item as a vector, represent a user as an aggregate of the vectors they’ve engaged with, and recommending becomes a nearest-neighbor query — exactly the operation you’ve already learned.

This reframing matters because it means you don’t need a dedicated ML team to ship useful recommendations. If you have item embeddings in a vector database for search, you’re one query away from “related articles,” “similar products,” and “customers also viewed” features. Embedding-based recommendations also sidestep the cold-start problem for new items: a product with zero interaction history still has a description, and a description still embeds, so it’s recommendable from day one.

Concretely, you’ll embed item content (titles, descriptions, transcripts) and store the vectors with metadata. “Similar items” is a top_k query using an item’s own vector, excluding itself. For personalization, average the embeddings of a user’s recent interactions into a taste vector and query with that, then filter out already-seen items with metadata filters. From there you can layer on business rules, freshness boosts, or an LLM to explain why something was recommended.