What is a Vector Database?

Traditional databases retrieve rows by exact keys or filtered conditions. A vector database instead stores each piece of content, such as a document chunk, an image, or a product description, as an embedding, a vector of numbers capturing its semantic meaning, and answers queries asking for the most similar items to a given one using approximate nearest-neighbor search across potentially millions of vectors, returning results ranked by semantic closeness rather than keyword overlap. This is the core infrastructure behind retrieval-augmented generation: rather than relying only on a language model's frozen training data, a system can embed a user's question, search a vector database of the organization's own documents for the most relevant passages, and hand those passages to the model as grounding context before it answers. Choosing a vector database involves tradeoffs between search speed, index size, recall accuracy, and how well it supports filtering by metadata alongside the similarity search.