Similarity search over embeddings

An embedding model turns a piece of text or an image into a vector of a few hundred to a few thousand numbers, and similar items end up close together. Finding the nearest neighbours exactly means comparing the query with every stored vector, which gives perfect recall but grows linearly with the data. Approximate indexes such as HNSW (a layered graph) or IVF (clusters) answer far faster by looking at only part of the data, and trade some recall for that speed.

To tell whether a system really handles it, measure recall and latency together on your own vectors, since either one alone is easy to improve. Check how much memory the index needs, how long it takes to build and whether it can take updates and deletes, and how it combines similarity with metadata filters: if filtering happens after the approximate search, a selective filter can return fewer results than you asked for. Hybrid search, which mixes vector similarity with keyword relevance, is a common extra requirement.