Pensieve

LinkedIn’s Pensieve platform moved expensive calculations out of the moment someone requests job recommendations. Its neural networks separately turn information about members and vacancies into compact lists of numbers, called embeddings, which ranking models can compare. Pensieve calculates and stores those representations ahead of requests, updating job embeddings as postings change. That lets several recommendation and search systems reuse learned information without running the full embedding network for every candidate match. The 2020 account explains both halves of the design: precomputation keeps ranking responsive, while processing updates helps prevent recommendations from relying on stale representations.