vector database

advancedwidespread

A database storing numerical vectors derived from texts and images, enabling searches based on semantic similarity rather than exact word matches.

In plain terms

Imagine a library where books are not arranged alphabetically, but by topic: similar contents stand next to each other, while very different ones are far apart. When you search for something, the librarian does not spell out the title but takes you to the appropriate shelf and removes the nearest volumes. A vector database is such a library, except the 'shelf location' is a long sequence of numbers.

This number sequence is called a vector or embedding. A separate AI model creates it for every text segment, image, or audio recording, compressing the meaning of the content into it. The number sequences for "dog" and "hound" come close to each other, even though they share no common letters.

A traditional database answers precise questions: what is the invoice number, who lives at this address. A vector database answers: 'what is most similar to this?' This is why a chatbot can retrieve exactly the paragraph from your company's multi-thousand-page regulations that answers your question.

In today's AI systems, it mostly serves as the "external memory" of language models. The model or an AI agent looks up in it up-to-date, proprietary or company knowledge that it did not originally learn.

How it works

Chunking: documents are cut into smaller, self-contained meaningful parts, such as paragraphs.
Embedding: an embedding model creates a number sequence from every detail, typically consisting of several hundred or a few thousand numbers that describe the detail's meaning.
Indexing: the database arranges number sequences so that similar items can be found quickly, without having to scan everything for every query.
Search: the same model creates a number sequence from your question, and the database returns the few details closest to it. This is mostly approximate search: very fast, but not guaranteed to bring the best result.
Usage: the language model or agent receives the results and formulates the answer based on them, often citing the source.

An everyday example

You type into your bank's or service provider's chat window, "I lost my card, what should I do?", and receive the relevant part of the card blocking information, even though you did not write the word "blocking". Behind the scenes, a vector database likely retrieved the text semantically closest to your question.

Who is affected

developerIf you build a chatbot, search engine, or agent based on corporate knowledge, you will likely solve retrieval using vector search, either with a dedicated vector database or with this capability in your existing database. → Code

entrepreneur, executiveThis determines whether your company's AI assistant answers from your own documents and where your data is stored meanwhile.

employeeThe workplace AI search engine can therefore find internal materials even if you do not remember the exact terms. → Work & HR

everyoneBehind recommendation systems, image search tools, and customer service chatbots, such search is increasingly at work.

A common misconception

MythA vector database understands text and answers questions on its own.

In realityIt only stores and retrieves based on similarity. The embedding model performs the 'understanding,' while the language model formulates the answer from the found details.

Limits and risks

Similar is not the same as correct: the system may return text that is on-topic but outdated or does not answer your question.
With precise data, such as item numbers, names or dates, it is often weaker than traditional keyword search, so the two are frequently combined.
Text converted into number sequences is also sensitive data: the original content can be partially inferred from it, so it must be protected accordingly.
If the embedding model is replaced, the entire dataset must be recalculated because the numerical outputs of different models are not comparable.

Similar, but not the same

relational database – SQL databases store in tables and search for exact matches, while vector databases search for semantic proximity. Today, many relational databases also support vector search.
embedding – The embedding itself is the number sequence that describes the content; the vector database is the system that stores and searches these.
RAG (retrieval-augmented generation) – RAG is the full method in which the model answers after search; the vector database can be only the search component of this, and RAG can also work with other search methods, such as keywords.

Related concepts

Broader

database, AI infrastructure, information retrieval

Narrower

vektorindex, approximate nearest neighbor search, hybrid search

On the edge

embedding, semantic search, RAG, agent memory, knowledge graph

A brief history

Similarity search is an old research field; a key milestone was when Facebook researchers published the open-source search library FAISS in 2017. As a standalone product category, vector databases became widely known in the years following the end-of-2022 launch of ChatGPT, when language models began to be mass-connected to their own documents.

Where things stand

By the end of 2026, vector search is a standard building block of corporate AI systems and has also been built into traditional databases and cloud services. Alongside dedicated products, you will therefore increasingly encounter it as an add-on capability of an existing database.

Status: 2026 November

Also known as

vector database · vector database · vector database · vector store · vector database · vector DB · vector store · vector search engine

Sources

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.