Performing Similarity Search

Similarity search is the mechanical core beneath semantic search: given a query vector, find the k stored vectors closest to it. “Close” is defined by a distance metric — cosine similarity (angle between vectors, the default for text), Euclidean distance (straight-line distance), or dot product — and the choice must match how your embedding model was trained. The naive approach, comparing the query against every vector, is called exact or brute-force k-NN; the scalable approach, approximate nearest neighbor (ANN), trades a sliver of accuracy for orders-of-magnitude speed.

You need this layer of understanding because vector databases expose it directly. Every query API asks you for a metric, a top_k, and often ANN tuning parameters, and misconfiguring them produces silently bad results — the query runs fine but returns the wrong neighbors. Knowing that cosine similarity ranges from -1 to 1, that normalized vectors make cosine and dot product equivalent, and that similarity scores are only comparable within one model saves you real debugging time.

Hands-on, you’ll compute similarity yourself first — a few lines of NumPy: normalize, then np.dot(query, matrix.T) — to demystify what databases do. Then you’ll run the same search through a vector store, passing top_k and metadata filters, and inspect the returned scores to set sensible relevance thresholds. This topic is the bridge to indexing, where you’ll see how ANN makes it fast.