Semantic Search

Search that finds documents by meaning using vector representations instead of relying on exact keyword matches.

Also known as: Vector Search, Semantic Retrieval, Embedding Search

In plain English

Semantic search finds items by meaning rather than exact words. It’s like asking a helpful librarian for documents on a topic instead of typing exact keywords. Different phrasings that mean the same thing will still return relevant results.

In practice

In products, semantic search means indexing content with embeddings and running vector similarity queries so users find relevant documents despite paraphrasing. It improves discovery and user satisfaction but requires a vector store, compute for embeddings, and monitoring for relevance, freshness and privacy.

Under the hood

Data (text, images, etc.) are converted into dense numeric embeddings, typically via transformer-based encoders. Vectors are stored in a vector database and queried with nearest-neighbour search (often approximate) using metrics like cosine or dot-product; top candidates are commonly re-ranked or filtered. Semantic search is frequently paired with retrieval-augmented generation to supply LLMs with relevant context.

Example

We should add semantic search to the knowledge base so people can find policies even when they use different wording.

•