Vector database

An ordinary database finds records by exact matches: this customer ID, orders after this date, products containing this word. A vector database finds records by similarity of meaning, so a search for "cancel my booking" can find a help article titled "How to change or call off an appointment".

It is a supporting part of AI systems, not an AI model itself. Vector databases are most often used behind semantic search and RAG, where they supply the passages an AI model reads before answering. Some are standalone products; many ordinary databases now offer vector search as an add-on.

How it works

First, content is split into chunks, such as paragraphs of a help article. Each chunk is passed through an embedding model, which turns it into an embedding: a long list of numbers, often hundreds or thousands long, positioned so that chunks with similar meanings get similar numbers. The database stores each embedding with a link back to the original text.

When a query arrives, it is turned into an embedding by the same model. The database then looks for the stored embeddings nearest to it, using a mathematical measure of closeness. Comparing the query with every item would be slow for millions of chunks, so vector databases use indexes that find near matches very quickly at the cost of occasionally missing the single best one.

Results are usually combined with ordinary filters, such as language, product category or the customer's account, and often with keyword search too, because pure similarity can miss exact terms like model numbers or names. Changing the embedding model later means re-embedding every document, because vectors from different models cannot be compared with each other.

Example

A travel agency has hundreds of pages of tour descriptions, visa notes and supplier terms. It splits them into paragraphs, embeds each one and stores them in a vector database. When a staff member asks the internal assistant "which tours suit someone who cannot walk far?", the database returns paragraphs about accessible coaches, short city walks and step-free hotels, even though none of them uses the word "walk far".

The agency checks quality by writing twenty real staff questions with the paragraphs that should come back. Most are found, but questions about a tour code return nothing useful, because codes have little meaning to an embedding model. Adding keyword search alongside the vector search fixes those. The agency also sets a simple rule that whenever a tour page is edited, its paragraphs are re-embedded overnight, so the assistant never quotes an old price or itinerary.

Why it matters for your business

You need a vector database only if you are building an AI feature that searches your own content by meaning, such as a help assistant or an internal knowledge tool, and the content is too large to paste into a prompt. For a few dozen pages, simpler setups often work fine. If a supplier proposes one, ask how documents are kept up to date, how quality is tested, and where the data is hosted.

Common mistakes

• Building a vector database for a small amount of content that would fit directly in a prompt.
• Forgetting to re-embed documents when they change, so the system keeps serving old text.
• Relying on similarity alone for queries that contain exact codes, names or numbers.
• Splitting documents into chunks so small that each one loses its context.

Questions people ask

Do I need a vector database for AI?

Not for most uses. Chat assistants, writing help and small document tasks need none. A vector database becomes useful when you build a tool that must search a large collection of your own content by meaning, typically for a support assistant or internal knowledge search.

What is the difference between a vector database and a normal database?

A normal database finds records by exact values, such as an ID or a date range, and is the right tool for orders and customers. A vector database finds items whose meaning is closest to a query. Many systems use both, and some normal databases now support vector search.

What is stored in a vector database?

Embeddings, which are lists of numbers representing the meaning of each chunk of content, plus a reference to the original text and often extra details like the source page, date or category. The original documents usually stay in their own system or alongside the vectors.