What is a vector database? Definition and business implications

It is a database specialised in storing and retrieving vectors, called embeddings. For a given query, it finds the content closest in meaning within a corpus, without requiring an exact word match. It is the typical search engine of a RAG system.

Each document is represented by an embedding, a string of numbers (typically 768 to 3,072 dimensions) that encodes its meaning. The query is also turned into an embedding, and the database looks for the closest vectors using a similarity measure. To scale to billions of vectors, it uses approximate algorithms such as HNSW or IVF, which respond in milliseconds.

Not always. Below a few million vectors, an extension such as pgvector on PostgreSQL or a multi-model database (MongoDB Atlas, Elasticsearch) is often enough, at a much lower cost. A specialised actor is justified only beyond that, depending on volume and team maturity.

Because when you handle confidential internal documents, the jurisdiction of hosting matters as much as technical performance. For a law firm indexing 80,000 acts, the options ranged from a managed service such as Pinecone (Frankfurt region since May 2026) or Weaviate Cloud to internal pgvector; the firm chose pgvector for sovereignty and cost control.

The billing structure. SaaS offerings generally combine a monthly minimum (Pinecone Standard: 50 dollars per month) with pay-as-you-go billing for storage, reads and writes, so cost depends as much on query volume as on stored volume. A benchmark at your target volume avoids unpleasant surprises on the bill.