Vector Search

Traditional keyword search asks: "Does this document contain the words in the query?" Vector search asks: "Does this document mean the same thing as the query?"

The process has three steps. First, the query is converted into an embedding — a numerical vector (array of numbers) that captures the semantic meaning of the text. Second, every document in the index is also represented as an embedding. Third, the search engine calculates the distance between the query vector and every document vector, returning the closest matches.

Distance is measured using metrics like cosine similarity or Euclidean distance. A query about "best plumber" and a document titled "experienced local plumber service" will have similar vectors even if they share no keywords. This is why vector search finds results by intent rather than by exact wording.

An embedding is a fixed-size vector (e.g., 768 or 1536 numbers) that represents the meaning of text. Modern embeddings are generated by neural networks trained on massive text corpora. They capture semantic relationships — "king" and "queen" have similar embeddings, as do "plumber" and "local home repair".

Because embeddings are numerical, the search engine can perform fast mathematical operations: comparing distances, clustering similar documents, or finding outliers. This makes vector search efficient even at scale. Embeddings have become the standard representation layer between language and search because they encode meaning in a form computers can manipulate quickly.

The quality of search results depends entirely on the quality of the embedding model. Better models (trained on more data, fine-tuned for your domain) produce better vectors and better search results.

Keyword search works on exact or near-exact term matching. A search for "NAP" returns documents containing the word NAP. If you search for "name address phone," keyword engines might return results about name tags and phone books.

Vector search captures meaning. The same query "name address phone" or "business identity consistency" returns documents about NAP — the concept — regardless of which exact words appear. This is particularly valuable for:

• Synonymy: "plumber" and "water service expert" match on intent, not words
• Short queries: Single-word searches benefit from meaning-based ranking
• Misspellings and typos: Embeddings capture intent despite surface variation
• Concept matching: A query about "local SEO rankings" matches documents about "geographic search visibility" even though words differ

The tradeoff: Keyword search is transparent (you can see why a result matched), while vector search is opaque (the similarity calculation happens in high-dimensional space humans can't visualize).

Vector search powers modern RAG (Retrieval-Augmented Generation) systems. When you prompt an AI agent with a question, RAG first retrieves relevant documents from a vector index, then passes those documents to an LLM to synthesize an answer.

This is how agents reference your documentation, your API endpoints, or your knowledge base without hallucinating. The workflow is: (1) embed the user's query; (2) search a vector database for the N most similar documents; (3) insert those documents into the LLM context; (4) LLM synthesizes an answer grounded in your actual data.

Vector search also underlies semantic similarity ranking in modern search engines. Google's understanding of results has shifted from keyword-based to embedding-based, particularly for complex queries and commercial intent. Perplexity and other AI search engines use vector retrieval as their core ranking mechanism. This makes understanding vector search essential for both search optimization and AI integration.