How should Vector search and similarity be designed for production?

Choose indexing and similarity settings for the embedding model, tune top-k and thresholds on evaluation data, and combine semantic relevance with metadata constraints when appropriate.