As AI becomes more accessible to the common person everyday, it also makes sense to implement a robust architecture that supports AI solutions. To that end, Microsoft has updated its .NET and JDBC ecosystems to natively support vector data types, ensuring high-performance environments for AI workloads.
When it comes to .NET, there is a new Sqlvector class in Microsoft.Data.SqlClient 6.1.0. This is capable of handling data in a vector format, rather than the previously used and inefficient JSON array format. Compared to that, the vector type improves performance by up to 50 times in read operations, 3.3x in write operations, and a 19x boost in bulk copy operations. These performance tests were conducted on SQL Server 2025 Preview with a vector column of max 1998 size, and 10,000 records for each
operation.
In addition, since JSON serialization is not required anymore, there is lower memory consumption too due to the absence of lengthy string representation. Finally, 32-bit floating-point vectors are natively supported right now but there is extensibility available for other numeric data types in the future too.
Switching over to the JDBC environment, we have a new VECTOR data type in the Microsoft JDBC driver for SQL Server 13.1.0. It can be directly used in insertions, select, stored procedures, and bulk copy. String-based vector handling has been removed in this version too, which makes it ideal for integration with Java-powered AI applications, including those offering semantic search.
It is important to note that these improvements are applicable on a wide variety of platforms, including SQL Server 2025 (17.x) Preview, Azure SQL Database, Azure SQL Managed Instance, and SQL database in Microsoft Fabric Preview. It is also worth highlighting that these improvements have been achieved through optimizations to the Tabular Data Stream (TDS) protocol, so they will only be visible on TDS version 7.4 or higher. Clients that don"t support this protocol will continue to use the varchar(max) data type and treat the data as a JSON array for backward compatibility, which means that they won"t see any of the aforementioned enhancements.