Vector database
Also known as: vector DB, vector store, vector search database
A vector database is a database built to store embeddings, the number lists AI models use to represent meaning, and to quickly find the ones closest to a query.
Instead of matching exact values like a normal database, it answers questions such as 'which documents are most similar to this question?' While vector databases became famous during the generative AI boom, many teams now find that adding vector search to a database they already run is enough.
Vector database, explained at three levels
At home
A normal database is like a filing cabinet sorted by name: you find a folder only if you know its label. A vector database is more like a librarian who can hand you the books most similar to the one you're holding, even with different titles. It stores each item as a list of numbers describing its meaning, and finds the lists closest to yours.
At work
A vector database stores embeddings, the number lists AI models use to represent meaning, along with the original content and labels such as date or department. Its main job is to answer 'what is most similar to this?' in milliseconds, even across millions of items. Businesses use it to power meaning-based search over documents, product recommendations, and chatbots that answer from company files.
At home
When you add a song, a document, or a photo, the system first turns it into an embedding and saves that list of numbers. When you search, your query becomes an embedding too. Checking every stored item would be slow, so the database builds an index, a shortcut structure that jumps quickly to the right neighbourhood. It then returns the handful of items whose numbers sit closest to your query, ranked from most to least similar.
At work
Vector databases rely on approximate nearest neighbour (ANN) indexes. HNSW builds a layered graph that a search hops through from coarse to fine, while IVF groups vectors into clusters and searches only the nearest clusters. Most products also store metadata so queries can filter by department, date, or permissions, and many support hybrid search that blends keyword (BM25) and vector scores. Options range from dedicated products such as Pinecone, Weaviate, Milvus, and Qdrant to pgvector inside PostgreSQL.
At home
Finding the exact closest match among millions of items means comparing against every one, which is too slow. Approximate indexes skip most comparisons by following a graph or checking only likely clusters, so they sometimes miss the true best match. Developers tune settings that trade speed against recall, and a recall of 95 to 99 percent is typical. While that sounds like a flaw, the speed gain is enormous. In my opinion, that's usually a fine trade, but worth knowing about.
At work
Architecturally, the choice is a dedicated vector store or vector search inside an existing database. HNSW indexes live largely in memory, so large collections need big servers, and rebuilding after bulk updates takes time. The biggest risk is access control: if permissions aren't filtered per document at query time, a chatbot can quote files a user isn't allowed to see. Data residency, backup, and cost at scale matter too, and recall should be measured against your own queries.
Real-world examples
At home
Aisha noticed her music app kept suggesting songs that 'felt' like her favourites, even from artists she'd never heard. The app was comparing song embeddings stored in a vector database.
At work
Thomas is a developer at a Calgary software company. Instead of adding a new vector database, he enabled pgvector in the PostgreSQL server his team already ran, keeping backups and permissions in one place.
How it compares
| feature | Vector database | Relational database |
|---|---|---|
| Query style | Find the most similar items | Find exact matches and filters |
| Data stored | Embeddings plus metadata | Rows and columns |
| Results | Ranked by similarity | Exact result set |
| Typical use | AI search, recommendations | Transactions, records |
Vector databases rank by similarity. Relational databases return exact matches.
For analysts
Before buying a dedicated vector database, I'd check whether your current database can do vector search well enough. One less system to secure and back up is a real saving.
History & origin
Researchers worked on nearest-neighbour search for decades, but exact search becomes slow with millions of high-dimensional vectors. Approximate methods made it practical: Facebook AI Research released the FAISS library in 2017, and the HNSW graph algorithm, published by Yury Malkov and Dmitry Yashunin in 2016, became the most popular index. Dedicated vector databases such as Milvus (2019), Pinecone, Weaviate, and Qdrant grew rapidly after ChatGPT's release in late 2022, as companies built chatbots over their own documents. At the same time, existing systems added vector features, including the pgvector extension for PostgreSQL, first released in 2021, plus Elasticsearch, MongoDB, and the major cloud databases. In my view, the vector database is becoming a feature more than a category.