Last reviewed: May 24, 2026

What is a vector database? Definition and business implications

A vector database is a database specialised in the storage and retrieval of vectors (embeddings). It allows, for a given query, finding the most semantically close content in a corpus, without exact lexical match. It is the typical search engine of a RAG system.

A vector database stores embeddings, that is, numerical representations (typically 768 to 3,072 dimensions) of the meaning of the indexed documents. When a query arrives, it is converted into an embedding, and the database searches for the closest vectors using a similarity measure (cosine, dot product). To scale, these databases use approximate algorithms (ANN, approximate nearest neighbors) such as HNSW or IVF, which sacrifice marginal precision for response times in milliseconds on billions of vectors. The market in 2026 includes three categories. Pure specialised actors: Pinecone, Weaviate, Qdrant, Milvus. Relational database extensions: pgvector for PostgreSQL, Azure SQL with extensions. Multi-model databases: MongoDB, Elasticsearch, Redis, which have integrated vector search. The choice depends on team maturity, volume, and the need to combine vector search with classic structured search.

Concrete example

A 150-employee law firm indexes 80,000 historical legal acts to allow its partners to query its own internal case law. Three technical options compared. Pinecone (proprietary cloud): 380 euros per month for this volume, setup in 2 days, US hosting. Qdrant Cloud Europe: 220 euros per month, setup in 3 days, Frankfurt hosting, GDPR-compliant. pgvector on internal PostgreSQL: near-zero cost if the DBMS already exists, setup in 7 to 10 days, full control. The firm chose pgvector for sovereignty and cost, despite the longer setup.

To ask your IT director

“Which vector database do we use for our RAG deployments, and do we really need a specialised actor?” This question, asked of your IT director or provider, often reveals costly over-engineering. Three sub-questions to chain. First, does the volume processed justify a specialised actor (Pinecone, Weaviate)? Below a few million vectors, pgvector or a multi-model extension (MongoDB Atlas, Elasticsearch) covers 90% of needs, at a much lower cost. Second, where is the database hosted? If you handle confidential internal documents, the jurisdiction of hosting matters as much as technical performance. Third, does the marginal cost increase linearly or in tiers? Several SaaS offerings bill per pod class, which creates sharp cost jumps at certain volumes. A benchmark at your target volume avoids billing surprises.

Sources

Efficient and robust approximate nearest neighbor search using HNSW graphs, Malkov & Yashunin, arXiv:1603.09320, 2016. https://arxiv.org/abs/1603.09320 (accessed 2026-05-24)
Vector database comparison, Pinecone documentation, 2026. https://www.pinecone.io/learn/vector-database/ (accessed 2026-05-24)

← Back to glossary

What is a vector database? Definition and business implications

Concrete example

See also

Further reading

Sources