What is a vector database?
A database that indexes data by meaning, not exact words: the memory RAG searches for content similar to a question.
A vector database stores and searches embeddings: numerical representations of the meaning of texts, images or other content. Two sentences saying the same thing in different words ("cancel the contract" and "early termination") produce nearby vectors, so search finds content similar in meaning, not just content sharing the same words. It is the infrastructure underneath semantic search and, above all, RAG: when an AI assistant must answer on company documents, the vector database is what locates the relevant chunks in milliseconds before the model even drafts a reply. In practice there are three viable paths: dedicated engines such as Pinecone, Qdrant, Weaviate or Milvus, extensions of a database you already run such as pgvector on PostgreSQL, or the vector capabilities now built into data platforms such as Databricks, Snowflake or BigQuery, each striking a different balance between operational simplicity and performance at large volume.
What it is for
It is the infrastructure of semantic search and above all of RAG: when an AI assistant must answer on company documents, the vector database is what finds the relevant chunks in milliseconds to pass to the model. The practical options are three: dedicated engines (Pinecone, Qdrant, Weaviate, Milvus), extensions of databases you already run (pgvector on PostgreSQL has become many projects' default), or the vector capabilities now built into data platforms (Databricks, Snowflake, BigQuery, managed cloud services).
The sensible choice
For most companies the question is not "which vector database to buy" but "do I really need one more component?". If you already run PostgreSQL or a modern data platform, the best answer is often to use the vector capabilities of what you own: one less piece of stack to manage and permissions already in place. A dedicated engine earns its keep at serious volume and latency requirements. And the usual order applies: first the quality of the indexed content, then the technology that searches it.
Related terms
- RAG · A technique that grounds an LLM in your company data: it retrieves relevant documents and feeds them to the model before it answers.
- Knowledge graph · A network representing company data as entities and relationships: customers, products, contracts and the links connecting them.
- Data lakehouse · A data architecture combining the flexibility of a data lake with the reliability of a data warehouse in one platform.
- Embeddings · A numerical representation of the meaning of a text or an image, used to compare content by similarity.
- Hybrid search and reranking · The ranking stage of RAG: it fuses lexical and vector search, then a reranker reorders the top candidates.
A term that hits close to home? Let's talk.
CONTACT ME