Semantic Vector Search

Keyword search matches text. Semantic search matches meaning. A keyword engine looks for the same characters you typed, so "puppy" never finds a page that only says "dog". A semantic engine compares what the words mean, so both land together.

How it works. An embedding model turns a piece of text into a vector, a list of numbers (often several hundred to a few thousand) that works like coordinates in a space of meaning. Texts with similar meanings end up close together. To compare two vectors, the usual measure is cosine similarity: the cosine of the angle between them, from 1 (same direction, same meaning) down through 0 (unrelated) to -1 (opposite).

A tiny example. Give "dog" the vector [0.9, 0.1], "puppy" [0.85, 0.2] and "invoice" [0.05, 0.95]. Dog and puppy have a cosine similarity of 0.99, so they point almost the same way. Dog and invoice score 0.16, so they are nearly unrelated. A search for "dog" would rank a page about puppies far above a page about invoices, with no shared keyword needed.

Where it is weak. Embeddings are poor at exact tokens: an error code like 0x3F, a product SKU, a person's name. Those need exact matching, which is why production systems often use hybrid search, combining keyword search (for example BM25) and vector search, then merging the two rankings. Quality also depends on the embedding model fitting your language and domain.

The scale problem. Comparing a query against millions of vectors one by one is too slow for an interactive search. That is the job of a vector database and its approximate nearest-neighbor index.