Dense Retrieval
Also called Embedding Retrieval
Finding documents by meaning, using embeddings, so "car trouble" can match "engine won’t start".
Think of it like
A helpful clerk who understands what you mean, not just the exact words you said.
Example
A search for "how do I cancel" surfaces the "ending your subscription" article even though no words overlap.
How it actually works
Dense retrieval encodes both query and documents into vectors and matches on similarity in that space, capturing synonyms and paraphrase. The strength is semantic recall; the weakness is that rare exact terms — a part number, a name — can get blurred away. That is why it is often paired with sparse methods in a hybrid setup.
For product teams
Makes search forgiving of how users phrase things; can miss precise codes and jargon on its own.
For engineers
Bi-encoder embeddings + ANN over a vector store; strong on semantics, weak on exact-token and long-tail matches.
Related
- Embeddings — Powered by embeddings.
- Sparse Retrieval — Contrasts with sparse, word-matching retrieval.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome