Beyond Embeddings: Practical Vector Search with DiskANN in SQL Server
Microsoft Developer focuses on building real-world semantic and hybrid search experiences directly inside SQL Server, using DiskANN for fast approximate nearest-neighbor (ANN) vector search.
Overview
Vector support in SQL Server enables semantic search scenarios, but embeddings alone are often not sufficient for production-quality search. This session shows how to keep data in SQL Server while still achieving scalable, high-performance vector search.
What DiskANN enables in SQL Server
- A practical approach to approximate nearest-neighbor (ANN) search for vectors.
- High-performance vector search designed to scale to large datasets.
- A way to run semantic similarity queries inside the database, without exporting data to an external vector store.
Storing and modeling vector data
The session demonstrates how to:
- Store vector data in SQL Server.
- Model tables that mix:
- Vector columns (for similarity search)
- Traditional relational columns (for structured attributes)
- Text columns (for keyword/full-text scenarios)
Indexing strategy with DiskANN
Through demos, the presenter covers how to:
- Choose an indexing strategy for vector search using DiskANN.
- Think about index selection in terms of performance and scalability for real workloads.
Query patterns: hybrid search in one place
A key theme is combining multiple retrieval signals in SQL Server queries:
- Vector similarity (semantic matching)
- Exact filters (metadata filtering on relational columns)
- Full-text search (keyword/text matching)
This hybrid approach is positioned as a way to build richer and more precise search experiences on your own data.
Takeaways
By the end of the session, viewers should be able to:
- Understand when and why to use DiskANN for vector search in SQL Server.
- Store, index, and query vectors alongside relational and text data.
- Apply vector search patterns to real-world scenarios using existing SQL Server data.