Stockertim Dictionary · Artificial intelligence · Technology
Vector database
Also called vector store, vector index
In short
A vector database stores embeddings alongside their text and quickly finds the stored items closest in meaning to a new question or passage.
- Advanced
- Technology
- Checked
In detail
Ordinary databases are built to find exact matches: a customer number, a date, a word. A vector database is built to find near matches in meaning. Each stored item holds an embedding, the original text or a pointer to it, and details such as source, date and access rights. A query is turned into an embedding and the database returns the stored items whose embeddings lie closest.
Comparing a query against every stored item would be slow at scale, so vector databases use indexes that find close matches very quickly at the cost of occasionally missing the exact nearest one. Most also allow filters, so a search can be limited to documents from one department, newer than a date, or visible to the person asking.
In retrieval-augmented generation the vector database is the memory the model draws on, so its contents decide the quality of answers. Outdated, duplicated or wrongly labelled documents lead straight to poor answers, and missing access filters can expose documents to people who should not see them. The collection needs an owner, a refresh process and the same security as the source systems.
Worked example
Kite & Co Accountants loads 4,000 policy and guidance documents into a vector database for an internal assistant.
- Each document is split into passages of about 300 words, giving roughly 4,000 x 6 = 24,000 passages.
- Each passage is stored with its embedding, source, date and the teams allowed to read it.
- A question from the payroll team is matched only against passages that team may read.
- Passages from superseded policies are removed each month.
Answers draw on current documents the asker is entitled to see.
In practice
Teams running assistants over company documents use vector databases with permission filters, record the source and date of every passage, and remove outdated material on a schedule.
Watch out
- A vector database finds similar meaning, not exact facts; check retrieved passages.
- Copying documents into a vector store can bypass their original access controls unless permissions are carried across.
Learn it properly
Official sources
- AI data security: best practices for securing data used to train and operate AI systems · Cybersecurity and Infrastructure Security Agency
- Top 10 for large language model applications · OWASP Foundation
Facts last checked on 3 November 2026.
Education, not advice: the Stockertim Dictionary explains what terms mean. It never recommends buying or selling anything.