Vector Database

In the past, if you searched for ‘dog’, you would only find entries containing the word ‘dog’. A vector database also finds ‘puppy’, ‘four-legged animal’ or ‘pet on a lead’ – because it understands that these terms are semantically closely related.

This is made possible by a special form of data: embedding. Texts, images or other content are translated by an AI model into long sequences of numbers – the vectors. Content with similar meaning is assigned similar vectors, meaning they lie close together in ‘vector space’. It is precisely this proximity that the database uses to calculate similarity.

What is an embedding?

An embedding is the bridge between human meaning and machine-readable numbers. An AI model converts content into a vector consisting of often hundreds or thousands of numbers, which encodes its meaning in a way that machines can compare. The key effect is that the spatial proximity of two vectors corresponds to their similarity in content. ‘King’ and ‘queen’ end up close to each other, whilst ‘king’ and ‘bicycle’ are far apart. A vector database stores these embeddings and is optimised to find the nearest stored vectors for a query vector – even with millions of entries, in a fraction of a second.

Vector database vs. traditional database

These two types of database are not mutually exclusive, but rather complement each other. Many applications use a traditional database for structured business data and a vector database for semantic search across texts and documents.

The most important use case: RAG

Vector databases form the backbone of Retrieval-Augmented Generation (RAG) – currently the most common way of combining a large language model with a company’s own knowledge. The problem that RAG solves: a language model only knows the knowledge it was trained on and tends to invent missing facts. If you integrate a vector database containing your own documents, a query works differently: first, the database searches the company’s knowledge base for the text passages most relevant to the content; then, the language model is provided with precisely these passages as context and formulates its response based on them. In this way, the AI responds on the basis of verified, up-to-date company information rather than from memory – which drastically reduces hallucinations.

A concrete example

Over the years, a company has accumulated thousands of pages of manuals, guidelines and support documents. Employees can hardly find what they’re looking for in them anymore. The solution: all documents are converted into embeddings and stored in a vector database. If an employee now asks the question “How do I handle a complaint about transport damage?”, the system does not search for the exact words, but for the sections with the most relevant content – even if the relevant document is titled “Handling goods damaged in transit”. An integrated language model summarises the passages found into a clear answer and cites the source. A confusing mountain of documents is transformed into a searchable, responsive knowledge system.

Well-known vector databases

The market is young but vibrant. Popular solutions include Pinecone, Weaviate, Qdrant, Chroma and Milvus; alongside these, traditional databases such as PostgreSQL are being upgraded with extensions (pgvector). Which solution is suitable depends on the volume of data, the operating model (cloud or self-hosted) and data protection requirements – the latter in particular is a key criterion when dealing with sensitive company data. The OpenAI documentation on embeddings provides an easy-to-understand introduction to the concept of embeddings, on which everything is based.

Context

The vector database is one of the unsung key technologies of the current wave of AI. It does not generate answers itself, but it makes the difference between an AI that spouts generalities and one that responds based on a company’s own knowledge. Wherever the aim is to make large volumes of unstructured content searchable and usable for AI – from internal knowledge search to intelligent product search in an online shop – it forms the foundation. Anyone seriously planning to develop AI applications based on their own data will find it almost impossible to do so without it.