Vector Search
Retrieval by similarity of dense vector representations rather than matching of keyword strings.
Vector search is a retrieval paradigm in which documents and queries are represented as dense numerical vectors — typically the output of a neural embedding model — and relevance is computed as geometric similarity between vectors rather than overlap of keywords. The paradigm emerged from advances in learned text and image embeddings and has become the default retrieval approach for cases where semantic similarity matters more than exact term match: question answering, recommendation, image search, and retrieval-augmented generation.
Mechanically, a vector search system has three parts: an encoder that turns documents and queries into vectors, an index that stores those vectors in a structure supporting fast approximate nearest neighbor lookup, and a similarity metric — most often cosine similarity or inner product — used to score candidate matches. The encoder and the index are independent choices: a system might embed with a transformer and index with HNSW, or embed with a vision model and index with IVF-PQ. The index choice is dominated by scale: at thousands of vectors, brute-force comparison is fine; at millions or billions, an approximate nearest neighbor index is required.
The advantage of vector search is the ability to surface documents that are semantically related to a query but do not share its exact terms — paraphrases, translations, related concepts, and visual similarity. The cost is that the embedding model is opaque: it is hard to debug why a given document is or is not retrieved, and the system inherits the biases of the encoder. Vector search also struggles with exact-match needs — proper nouns, identifiers, and rare terms — which is why it is usually paired with lexical search in a hybrid pipeline. Index size and update latency are also operational constraints that grow with corpus scale.
Open questions include how to make vector search more interpretable to developers and end users, and how to combine dense vectors with sparse or learned-sparse representations to cover the cases that pure dense retrieval misses. There is also ongoing debate about the right dimensionality: higher-dimensional embeddings capture more nuance but cost more to index and search. As embedding models continue to improve, the question of whether vector search will subsume keyword search — or remain a complement to it — remains open.