How Vector Search Works in SQL Server 2025

SQL Server 2025 vector search technology can help developers add natural language search capabilities to existing SQL Server text data without requiring an external AI solution or copying data. Vector search technology is built into the SQL Server database engine and uses new T-SQL commands and functions that are discussed below.

Configure the Database

The first step is to configure the database to store vector data, which are arrays of numbers also called embeddings.

Vector data is stored in a new vector() data type, which is a numeric field optimized for array data. Vector columns can be added to an existing table, placing it close to the text data, or created in a separate table. The developer should decide on a schema that provides the best responsiveness for the application.

Connect to External LLM

Developers choose and connect to an external large language model (LLM). Although they can choose any LLM, running locally or in the cloud, Microsoft recommends LLMs offered by OpenAI, the Azure OpenAI service, or Ollama. Local LLMs provide a higher level of data security and control, because the data never leaves the customer’s environment, but it can suffer from slower performance and require ongoing maintenance.