This example demonstrates a minimal similarity search over precomputed embeddings, returning the top ten most semantically similar records to a query phrase. It uses DuckDB’s Expression API to construct a `list_cosine_similarity` function call that compares each stored embedding against the query’s OpenAI-generated embedding vector. The code chains relational operations—selecting all columns plus the computed similarity, sorting by descending similarity, and limiting to ten rows—before excluding the embedding and similarity fields for output. The data comes from the `rms_embeddings_01` table in an embedded DuckDB database file.