Definition
A vector database is a database specialized for storing large numbers of embeddings and efficiently searching them for the closest matches to a given query vector, based on semantic similarity rather than exact keyword match. It is the storage and search infrastructure that Retrieval-Augmented Generation relies on to find relevant source content quickly.
Key facts
- Specialized specifically for storing and searching embeddings (vectors), not general-purpose relational or document data.
- Finds matches by semantic similarity (mathematical closeness in vector space), not exact keyword or field matching.
- The infrastructure layer Retrieval-Augmented Generation relies on to quickly locate relevant source passages at scale.
- Performance and relevance of results depend on the quality of the embeddings stored, not only the database's search algorithm.
How does Vector database work in maintenance?
A vector database stores large numbers of embeddings (vector representations of documents, passages, or other content) and, given a new query vector, efficiently finds the closest matches by measuring mathematical similarity across potentially millions of stored vectors, far faster than comparing against each one individually.
What does Vector database require?
Embeddings are the vector representations a vector database stores. Retrieval-Augmented Generation relies on a vector database to store embeddings of source documents in advance, then, when a question arrives, converts the question into an embedding and searches the vector database for the closest matching source content to feed into the generative step.
FAQ
Frequently asked questions
How are vectors stored and searched by similarity?
Through a vector database: a database specialized for storing embeddings and efficiently finding the closest matches to a query by semantic similarity rather than exact keyword matching.
Last reviewed: