Back to blog
AI•June 4, 2026•4 min read

Embeddings made our search understand what people meant

Keyword search has a hard ceiling: it matches the words you typed, not the thing you meant. Across a 185k-card catalog, that gap is huge - collectors describe a card by vibe, by partial memory, by the wrong name. The fix is embeddings: turn every catalog entry and every query into a vector, and "similar meaning" becomes "close in space." Suddenly search finds what someone meant, not just what they spelled.

Meaning, not string matching

An embedding model maps text (or images) into a space where related things land near each other. Pick one that fits your domain and "the holographic charizard from base set" lands next to the right card even if the exact words never match a database column. That's the same leap that separates a frustrating search box from one that feels like it gets you - and it's plumbing, not magic: get the embeddings and the retrieval right and the experience transforms.

Keyword search asks "do these strings match?" Embeddings ask "do these mean the same thing?" Users only ever wanted the second one.

The same machinery powers real RAG

This isn't a niche search trick-it's the retrieval half of retrieval-augmented generation. When you ground an LLM in your own data, the quality of the answer is mostly the quality of what you retrieved, and embeddings are how you retrieve by meaning. Most "the AI gave a bad answer" bugs are really "we embedded and retrieved the wrong context" bugs. Invest there.

How to get embeddings working for you:

  • Embed both sides - catalog entries and the query - into the same space.
  • Test on your domain's own queries so "close in space" means "relevant to your users."
  • Treat retrieval quality as the metric: a great LLM over the wrong context still fails.
  • Reuse it for RAG: the search that understands intent is the retriever that grounds your model.