Vector Search for Products

A vector search system has three pieces. The first is an embedding model that turns text (or images) into vectors of a few hundred dimensions. The second is a vector index that stores those vectors and supports nearest-neighbour search across millions of items in milliseconds. The third is the query layer that embeds the user's text and ranks products by similarity.

For Shopify catalogs, the index is built once and updated as products change. Each product is usually represented by multiple chunks: title, description, attributes, sometimes review snippets. When the shopper searches "warm jacket for Delhi winters under ₹4,000," the embedding model maps that query into the same space, and the index returns jackets with similar embeddings, even if those jackets do not contain the exact words "Delhi" or "winter."

A modern setup blends vector and keyword search. Pure vector retrieves on meaning but can miss exact-match SKUs. Hybrid retrieval (BM25 plus vector) combines both, and a cross-encoder re-ranker can sharpen the final order. Filters for stock, price, and category sit on top so the result respects business rules.

For example, a shopper searches "gift for my dad who likes fishing." Keyword search returns nothing because the catalog has no products tagged "dad" or "fishing." Vector search returns fishing rods, tackle boxes, and a fishing-themed mug because those products embed close to the query. A second example: "comfortable shoes for standing all day" returns nursing shoes, work boots, and cushioned trainers, none of which match the literal query terms.

For Shopify merchants, the on-site search box is one of the highest-converting surfaces in the funnel. Shoppers who use search are typically high-intent. A search that returns "no results" or irrelevant products on a query the catalog can actually serve is a direct revenue leak. Vector search closes most of that gap by understanding the shopper's meaning rather than only matching their words.

The second benefit is downstream: once products are embedded, the same vectors power AI recommendations, related-products blocks, and chatbot retrieval. The merchant pays the embedding cost once and reuses it. Shop Me's shopping assistant uses this pattern, so every product the assistant recommends is the result of vector matching against the live catalog.