What it means
A vector database holds embeddings: numeric representations of text where similar meanings sit close together in space. When a user asks 'how do I cancel?', the AI agent converts the question to a vector and asks the database for the closest stored passages. The match works on meaning, not on the word 'cancel'.
The most common vector databases in production today are Pinecone, Weaviate, Qdrant, and the vector extensions inside PostgreSQL (pgvector). For small knowledge bases (under 10,000 documents), the choice rarely matters. For large or multilingual knowledge bases, the engine matters more.
Why it matters
Vector databases are what make retrieval-augmented generation (RAG) work. Without them, an AI agent either hallucinates answers about your business or is limited to whatever fits in its context window. With them, it can ground every answer in the right passage of your actual knowledge base.
They are also what makes AI multilingual without translation. A customer asks in Bahasa, the agent finds the relevant English passage in your knowledge base (because semantic meaning is language-agnostic at the vector level), and answers back in Bahasa. No translation step needed.
Example
A furniture retailer's AI sales agent uses a vector database with 1,800 product passages: dimensions, materials, care instructions, warranty notes, common questions. A customer asks 'is this couch okay for cats?'. The agent finds the right passage (about stain resistance and fabric durability) and answers correctly, even though the word 'cats' was never in the source text.