Vector Database
Subject: Computer and Technology noun /ˈvɛk.tər ˈdeɪ.təˌbeɪs/ 1 definition
- database-systems
- information-retrieval
- machine-learning
- data-management
-
Definition (noun)
Entry #254 · UpdatedA Vector Database is a specialized data management system built to store, index, and query high-dimensional numerical vectors—known as embeddings—alongside structured metadata. Unlike traditional relational databases that rely on exact keyword or range matches, it uses mathematical distance metrics such as cosine similarity, Euclidean distance, or dot product to find semantically similar items. By pairing approximate nearest neighbor indexing with real-time updates and metadata filtering, it enables low-latency similarity search across massive unstructured datasets for applications like retrieval-augmented generation and recommendation systems.
Common Types
Native (Purpose-Built) Vector Databases: Standalone systems designed specifically for vector storage, distributed sharding, and approximate nearest neighbor indexing, optimized for large-scale, low-latency similarity search in dedicated AI pipelines.
Database Vector Extensions: Index modules and vector column types added to established relational, document, or search engines—such as
pgvectorfor PostgreSQL—enabling hybrid vector and structured queries inside a single transactional system.Embedded Vector Databases: Lightweight, in-process engines that store vector indexes in local files or object storage directly within an application, ideal for edge deployments, local development, and single-node workloads without cluster overhead.
Serverless Vector Databases: Cloud-managed services that decouple stateless query compute from durable object storage, automatically scaling capacity and index compaction on demand for bursty production workloads.
Example
An e-commerce support team embeds 500,000 product manuals and resolved helpdesk tickets into 1,536-dimensional vectors and stores them in a vector database tagged with product category and firmware version. When a customer asks why their smart thermostat fails to pair over Wi-Fi, the application embeds the question and runs a filtered similarity query (
category = "thermostat") to retrieve the top five relevant troubleshooting passages in under 20 milliseconds.