What is Vector Database?
Also called Vector Store, Vector Search Engine.
A vector database is a data store built to hold high dimensional numeric vectors and to find the ones most similar to a query vector quickly. It indexes embeddings produced by machine learning models and supports nearest neighbor search, usually with metadata filters. Such systems underpin semantic search and retrieval for language model applications.
Traditional databases match exact values or text tokens. A vector database instead measures geometric closeness between numeric arrays, typically hundreds to a few thousand numbers long. Because comparing a query against every stored vector becomes slow at scale, these systems use approximate nearest neighbor indexes that trade a small amount of accuracy for large gains in speed.
Common index structures include hierarchical navigable small world graphs, inverted file indexes built on clustering, and quantization schemes that compress vectors into smaller codes. Each exposes tuning parameters that balance recall, latency, and memory. Raising the number of candidates examined per query generally improves recall and increases response time.
Most deployments also store metadata beside each vector, such as document identifier, author, timestamp, or access scope. Filtering on that metadata during search keeps results within a tenant, a date range, or a permission boundary. Handling filters and vector similarity together is a distinguishing engineering problem, since filtering naively after search can return too few results.
Vector capability is available both as dedicated services and as extensions to established relational and search engines. The practical choice often turns on operational preference, since teams already running a general purpose database may prefer an extension over a separate system. Scale, filter complexity, and update frequency also weigh on the decision.
Key points
- Stores embeddings and finds nearest neighbors by distance.
- Approximate indexes trade slight accuracy for large speed gains.
- Metadata filters keep results within tenants or date ranges.
- Available as dedicated services or as database extensions.
In practice
A company indexes twelve thousand help articles. Each article section is embedded and stored with fields for product area and last updated date. When a user asks about invoice exports, the query vector is compared against the index, filtered to the billing product area, and the ten closest sections return in a few tens of milliseconds for the model to read.