Embedding

An embedding model turns a piece of content into a vector, a fixed-length list of numbers. Content with similar meaning produces vectors that point in similar directions, even when the wording differs, so how do I get my money back and refund policy end up close together. Comparing vectors gives you a measure of semantic similarity.

Embeddings power semantic search, retrieval for RAG, duplicate detection, clustering, recommendations, and semantic caching. The typical setup embeds your content once at ingestion, stores the vectors in a vector index, embeds each query at request time, and looks up the nearest stored vectors.

A few decisions matter. Vectors from different models aren't compatible, so changing models means re-embedding everything. Higher dimensions cost more storage and memory, though some models let you shorten vectors with a modest quality trade-off. Some models expect different settings for queries and documents. And leaderboard rankings don't transfer reliably to your domain, so test candidates on your own data.

Embeddings are also weak on exact matches: product codes, error IDs, and rare names often retrieve poorly, which is why hybrid search adds keyword matching alongside vectors.