What a vector database is and when you really need one

Once your documents have become embeddings, you need somewhere to store them and query them. That is what a vector database does: it receives the question turned into numbers and returns the pieces of document that are the closest match.

There are dedicated products and extensions to traditional databases. For many small and medium-sized businesses (SMBs) the second route is enough: if you already have a database, you can often add vector search without a new system to maintain.

An example in an SMB

A payroll and employment consultancy has years of circulars and internal opinions. It splits them into paragraphs, creates the embeddings and saves them in a vector database. When a colleague asks a question, the internal assistant retrieves the right paragraphs and answers by quoting them.

The most common mistakes

• Choosing a complex tool for a few thousand documents, when a simple solution works perfectly well.
• Not storing the original text and the source together with the vector: you then can't tell where the answer comes from.
• Neglecting permissions: if a document is confidential, it must not end up in everyone's answers.
• Splitting documents into pieces that are too short or too long without testing: this choice has a big effect on what the assistant manages to find.

Why it matters to you

The vector database is the part that stays switched on every day once the assistant is in production. Who hosts it, where the data sits, who does the backups and how you delete a document that must no longer appear are questions to put to the supplier before you start, not after.

All the words of AI, explained simply, are in the glossary.