SQL Server introduces a native vector data type and vector indexing for embedding-based search. How should you generate and store embeddings-and what additional signals should you include-to maximize retrieval quality from a vector store?

This session draws on a real-world implementation with more than 1.3 million documents indexed in SQL Vector index. I cover a scalable, structure-aware approach to embedding document content, including how to represent document hierarchy and how to incorporate extracted keywords into the index to influence ranking. Once the embedding and indexing strategy is in place, I explain the core logic behind approximate nearest neighbor (ANN) algorithms and what to expect from vector search at scale.

Finally, after retrieval, we look at options for enriching a RAG pipeline with relevant adjacent data (for example, surrounding sections, metadata, or related entities). The approach leverages the original document structure so the model receives the right context-not just isolated chunks.