At Pinterest, the mission is to inspire users to create a life they love, and finding the right content online is crucial to this mission. The company's recommendation system involves multiple stages, including retrieval and ranking, to serve the right content to users. The ranking model is a powerful transformer-based model that captures users' long and short-term engagement, but the retrieval system was previously based on heuristic approaches. To improve this, Pinterest built an internal embedding-based retrieval system learned from logged user engagement events, which has been deployed for homefeed and notification. The system uses a two-tower approach, where one tower learns the query embedding and one tower learns the item embedding, allowing for efficient online serving with nearest neighbor search. The model is trained using a sampled softmax approach to correct for popularity bias, and user long-term engagement, profile, and context are encoded as input. The system design involves breaking down the item embeddings into online serving and offline indexing, with an auto-retraining workflow to refresh the learned knowledge of users and capture recent trends. To ensure model version synchronization, a piece of model version metadata is attached to each ANN search service host, which contains a mapping from model name to the latest model version. The learned retrieval candidate generator has achieved top user coverage and top three save rates, and has helped deprecate two other candidate generators with huge overall site engagement wins.