Vector searches are usually evaluated for speed and if they are 'good enough' for production. This session examines vector search in SQL Server 2025 as an accuracy engineering problem, treating quality, latency, and scalability as explicit design variables rather than implicit side effects.

We start by understanding the basics of vector search in SQL 2025, and proceed to look into distinguishing exact nearest‑neighbor (ENN) search from approximate nearest‑neighbor (ANN) search.DiskANN is necessary for performance, but those gains come at the cost of approximation errors. Understanding ENN and differentiating it from ANN, with pros and cons, forms the next part of the talk - this is demo-heavy, with solid examples of how these two techniques differ. We then get into understanding 'recall' as an engineering metric that helps quantify ANNs and understand trade-offs in using them. Practical techniques for building baselines using ENN and evaluation queries, including how to interpret 'recall' will be included for user‑visible relevance rather than abstract mathematical similarity. We will also examine latency distributions, tail behavior, and non-linear trade-offs, and include dataset size, vector dimensionality, top-k selection, and predicate selectivity as parameters for varied outcomes and wins versus losses.

User takeaways towards the end will include a clear understanding of how to use vector search as a measurable, tunable, subsystem governed by explicit quality and performance criteria. This will help make engineering decisions based on real data rather than intuition.The learning will help anyone dealing with vector search queries, although demos are specific to SQL Server 2025.