Semantic Search

The VECTOR type enables Approximate Nearest Neighbor search entirely inside MariaDB. Semantic retrieval can find relevant information even when the query wording differs from the original documents. Vector search can return conceptually related content even when it does not contain the exact query string. Vectorization is not intended to perform keyword matching. MariaDB's <=> operator computes cosine distance between two vectors. A smaller cosine distance result means two vectors are closer and more similar. Vectorization enables search based on context and intent. Similarity search can rank table vectors by their cosine distance from a target vector. Cosine distance is suited to embeddings because it compares vector direction rather than magnitude. Similarity searches work by finding stored vectors near the query vector. Semantic search is described as the most direct application of vector databases.