What is vector search?
Vector search is a way to search by meaning, where texts are described as numbers so the search finds what means roughly the same as the question.
Ordinary search looks for the same words as in the question. Vector search looks for the same meaning. Someone searching for holiday request can then find a passage about annual leave, even if the words holiday request never appear.
Each passage of text is turned into a vector, a long row of numbers that describes what the text is about. Texts with similar content get similar numbers, and the search compares them with the question.
It is also called semantic search. The technique is often used when AI answers from your own documents. It decides which passages reach the model and so how good the answer becomes.
How it works
A dedicated model reads each passage and gives it a position in a space with a great many dimensions. Passages about the same thing end up close together. A question is placed in the same way, and the search fetches the passages that sit nearest to it.
Why it matters
Staff rarely phrase things the way documents do. Vector search finds the answer anyway, even when the question uses everyday words and the procedure uses formal ones. It is often what decides whether an AI assistant feels accurate.
What to keep in mind
The search finds what resembles the question. That does not always mean the passage answers it. A passage about last year's holiday rules can sit just as close as the one that applies this year.
Passage size, headings and versions therefore matter a great deal. Combine it with ordinary keyword search, and always show the citation so users can see what was retrieved.
What it means for you
The search needs to respect the same permissions as your documents. Otherwise the wrong people can see the wrong content.
Common questions
How is it different from ordinary search?
Ordinary search looks for the same words. Vector search looks for the same meaning, so a question about holidays also finds a procedure on annual leave.
Is keyword search still needed?
Often, yes. Product names, section numbers and abbreviations are found most reliably with exact words. Many solutions combine both.
Could vector search show documents a user is not allowed to see?
It should not. Search needs to follow the same access control as your other systems, so each person only gets results from what they are allowed to read.
Related terms
Read more
All terms Updated