In context
Traditional search is keyword-matched. A shopper asks for "something cosy for a rainy day" and a keyword index has no idea where to look. Embedding-based search converts that query into the same vector space as your products, then returns the products whose vectors sit closest to the query.
The "closeness" metric is usually cosine similarity. Most modern stores use 384, 768, or 1536-dimension vectors generated by a model like text-embedding-3-small. Higher dimensions capture more nuance at the cost of storage and query speed.
Embeddings power more than search — they're also how AI does product clustering, "similar items" recommendations, duplicate detection, and zero-shot categorization.
How Zubby uses this
Zubby embeds every product, variant, and policy document into a tenant-isolated pgvector index. We use a 1536-dimension OpenAI embedding model and re-embed any row that changes via webhook — so the agent never serves stale meaning.
We pair vector similarity with structured filters (price, stock, collection) and a re-ranking pass so results respect both semantic intent and your business rules.