Activity embeddings and new curation tools
Practitioners recommend using activity embeddings to capture engagement patterns and match similar users via vector databases for scalable feed personalization, and BetaList highlighted Recomate as an AI tool for tailored picks on shows, food and activities. The Social briefing aggregated the technical note on embeddings and the product spotlight on Recomate as current signals in recommendation tooling ( ).
Recommendation software is shifting from hand-built rules to “activity embeddings,” where a person’s clicks, watches, saves, and skips are turned into coordinates a machine can compare at scale. (developers.google.com) Google’s recommendation systems course describes a common pipeline with candidate generation, scoring, and re-ranking, and says embeddings are used to represent items and queries inside that pipeline. (developers.google.com) A vector database is the storage layer for those coordinates: it keeps the numeric representations and returns the nearest matches fast enough for live feeds, search, and recommendations. OpenAI’s Pinecone cookbook says vector databases are used to store, manage, and search embedding vectors for production systems, including recommendation services. (developers.openai.com) Google Cloud’s Vertex AI Vector Search documentation says the service is built for search and recommendation systems and is based on ScaNN, a nearest-neighbor method from Google Research. The same page says a minimal setup can run at under $100 a month for moderate throughput, which helps explain why smaller teams are now testing infrastructure that used to be reserved for the largest platforms. (cloud.google.com) That technical pattern is showing up alongside consumer-facing curation tools. BetaList’s listing for Recomate says the iOS app learns a user’s tastes from a short questionnaire and feedback, then recommends television, movies, restaurants, bars, and activities based on mood and location. (betalist.com) BetaList also says Recomate adjusts its suggestions using time, weather, and nearby events, and sends alerts when new options fit the user’s profile. That is the same basic recommendation problem as a feed, applied to leisure choices instead of posts or products. (betalist.com) The broader tooling stack is also getting easier to buy rather than build. NVIDIA says its Merlin framework is designed for recommender systems at scale and covers retrieval, filtering, scoring, ordering, training, and deployment across hundreds of terabytes of data. (developer.nvidia.com) In practice, that means recommendation teams can mix off-the-shelf embedding models, managed vector search, and packaged recommender frameworks instead of wiring every component from scratch. The result is a market where the same core machinery can power a social feed, a movie picker, or a restaurant suggestion app. (developer.nvidia.com; cloud.google.com; developers.openai.com)
