What is a vector database?

Vector database: definition

Classic databases search for exact values: customer number, date, keyword. A vector database answers a different question: which stored content is most similar in meaning to a query? To do this, it stores embeddings and compares the distances between the vectors. The method is called similarity search or semantic search and also finds hits in which the query’s keywords do not appear at all.

Besides the vector, it usually stores the original text or a reference to it, plus additional information called metadata, such as document type, language, department or access right. This lets searches be narrowed to specific areas and permissions be respected. A vector database therefore does not replace a master-data database but adds a search by meaning.

How similarity search works

For a query, the question is converted into a vector. The database looks for the nearest neighbors. To do this in fractions of a second even with millions of entries, it uses special indexes (for example HNSW) that search only approximately and are very fast as a result.

The choices range from specialized products such as Qdrant or Weaviate to extensions of existing databases such as pgvector for PostgreSQL. For many mid-sized applications the extension is enough: it avoids an additional system and can be backed up together with the rest of the data.

Real-world example: knowledge system for customer service

A manufacturer of building technology wants its customer service to answer questions about devices, faults and spare parts faster. Manuals, service bulletins and approved solutions are split into sections, converted into vectors and stored in a vector database. An AI assistant retrieves the relevant passages for each question and summarizes them with source references.

Through the metadata, field service only gets approved documents; internal drafts stay excluded. When a manual is updated, only the affected part is recalculated, so the knowledge base stays current. The assistant’s answers point to the underlying document so employees can quickly double-check them.

What to watch for

A vector database is only as good as its content and how that content is split. Sections that are too large dilute the hits; sections that are too small lose context. Permissions must apply at search time, not only at display. Where personal data is stored, server location also matters.

Nano Concepts plans and builds such knowledge systems for mid-sized companies, from document import to securing access. First we check whether your use case needs vector search at all. On request we run the solution in a European hosting environment and connect it to your document management system, ticket system or website.

• Check whether a vector extension of your existing database is enough
• Plan metadata and permissions from the start
• Plan for updating and deleting documents
• Clarify EU hosting and encryption
• Measure result quality regularly with test questions