Vector Search

Vector search is a retrieval method that represents queries and content as numeric vectors, called embeddings, and returns the items whose vectors are closest to the query's vector. Because embeddings place texts with similar meaning near each other, vector search can match 'automotive servicing' to a page about 'car maintenance' even when they share no words. It is the semantic half of most modern retrieval pipelines, including those behind RAG systems and AI answer engines.

The mechanics have three parts. First, an embedding model converts every chunk of content into a vector, typically with hundreds to a few thousand dimensions; Google's current Gemini embedding models default to 3,072. Second, the vectors are stored in an index. Third, at query time the query is embedded with the same model and the index returns the nearest neighbors by a distance measure such as cosine similarity, dot product or Euclidean distance.

Comparing a query against every stored vector is too slow at scale, so production systems use approximate nearest neighbor (ANN) indexes that trade a little accuracy for large speed gains. The most widely used is HNSW (Hierarchical Navigable Small World), described by Yu. A. Malkov and D. A. Yashunin in 2016: a layered graph in which a search starts at a sparse top layer and descends toward close neighbors, giving roughly logarithmic search time (Malkov and Yashunin). Another common family, IVF, partitions vectors into clusters and searches only the closest clusters.

Vector search ships in dedicated vector databases and in general-purpose systems. The Postgres extension pgvector, at version 0.8.7 as of October 1, 2026, supports both HNSW and IVFFlat indexes and has offered HNSW since version 0.5.0 in August 2023. Elasticsearch provides a kNN retriever, and MongoDB Atlas offers vector search with built-in reranking.

Vector search has known weaknesses. Embeddings blur exact identifiers, so a query for a specific error code or part number may return pages about similar codes. Pure vector stores have also been criticized for weak filtering: Perplexity wrote in 2025 that it saw 'vector databases that couldn't adequately support even simple boolean filters'. That is why many production systems, Perplexity's among them, run vector and keyword retrieval side by side as hybrid search and finish with reranking.

Vector search and semantic search are often used as synonyms. A useful distinction: vector search is the mechanism (nearest-neighbor lookup over embeddings), while semantic search is the broader goal of matching meaning and intent, which search engines pursued with other techniques long before vector databases existed.

For GEO, vector search rewards passages whose meaning is explicit. A passage that says plainly what a product is, who it is for and what problem it solves sits close to the many ways people phrase that need. A passage full of slogans, or one whose meaning depends on earlier paragraphs, produces a vague embedding that sits near nothing in particular. Because each chunk is embedded separately, every section should name its subject and make its point on its own.