Semantic search

Semantic search finds results that mean the same thing as the query even when they share no words. A user typing can't log in after changing my email can find an article titled Updating your account address. Under the hood, content is embedded ahead of time, the query is embedded at search time, and the nearest vectors become the results.

It is a feature in its own right, in help centers, document libraries, and product catalogs, and it's the retrieval step at the heart of most RAG systems. Building it well involves more than the vector query: chunking content sensibly, filtering by permissions and metadata, returning useful snippets, and handling the case where nothing relevant exists.

Its weakness is precision on literal terms. Product codes, error messages, names, and short exact phrases often retrieve poorly, because embeddings capture general meaning rather than specific strings. Pure semantic search can also return something vaguely related with confidence when nothing truly matches. That's why most production systems combine it with keyword search and set a similarity floor for a no-results state.

Evaluate it with real queries from your logs and labeled relevant results, not by eyeballing a few searches.