Dense Retrieval
An information-retrieval approach that encodes queries and documents as learned fixed-dimensional vectors and retrieves candidates by vector proximity.
Dense retrieval does not require a query and document to share the same surface words. If an encoder places them near one another in embedding space, semantic similarity can produce a candidate. The same property can create confident but irrelevant neighbors when the embedding model is weak for the domain.
Production cost is not only the encoder. Vector-index structure, the recall/latency trade-off, and update behavior determine end-to-end performance. FAISS vector-search index architecture focuses on that retrieval layer.