Vector database

A vector database is a system designed to store and search embeddings — fixed-length arrays of numbers that encode the meaning of a piece of text, an image, or a chunk of code. Where a keyword index asks "which records contain this word?", a vector database asks "which records mean roughly the same thing?" It answers by measuring the distance between vectors and returning the closest matches.

Under the hood, the database builds an approximate nearest-neighbor (ANN) index — structures like HNSW graphs or IVF clusters — so it can scan millions of high-dimensional vectors in milliseconds instead of comparing the query to every record one by one. You hand it a query embedding and a value for k, and it returns the top-k most similar items along with their similarity scores, typically using cosine similarity or dot product as the distance measure.

This matters because it is the retrieval half of retrieval-augmented generation. When a user asks an agent a question, the system embeds the question, looks up the most relevant passages in the vector database, and feeds them into the prompt as grounding. The same mechanism powers agent memory: past conversations and learned facts are saved as vectors and recalled on demand. For example, a support agent embeds an incoming ticket, retrieves the three closest help-center articles, and answers from those exact sources — accurate, current, and traceable instead of guessed.