One-Sentence Definition
A vector database is a specialized database designed to store, index, and search high-dimensional vectors -- the numerical representations (embeddings) that AI models use to capture the meaning of text, images, and other data.
How It Works
Traditional databases search by exact matches or predefined filters: find all users where country equals "France" or all products under ten dollars. Vector databases solve a different problem: find the items most similar to a given query based on meaning, not keywords.
When an AI application processes a document, image, or audio clip, an embedding model converts it into a high-dimensional vector -- an array of hundreds or thousands of floating-point numbers. The vector database stores this vector alongside the original content or a reference to it. At query time, the system converts the search query into a vector using the same embedding model and finds the stored vectors closest to it in the high-dimensional space.
The core technical challenge is speed. Comparing a query vector against every stored vector (brute-force search) is too slow when the database contains millions or billions of entries. Vector databases use approximate nearest neighbor (ANN) algorithms to make this fast. Common approaches include HNSW (Hierarchical Navigable Small World graphs), IVF (Inverted File Index), and product quantization. These algorithms trade a small amount of accuracy for dramatic speed improvements -- typically returning results in milliseconds even over billion-scale datasets.
The major purpose-built vector databases in 2026 include Pinecone (fully managed cloud service), Weaviate (open source with hybrid search), Chroma (lightweight, popular for prototyping), Qdrant (open source with Rust-based performance), and Milvus (designed for billion-scale workloads). Traditional databases have also added vector capabilities: PostgreSQL has pgvector, Elasticsearch added dense vector search, and Redis offers vector similarity search.
Why It Matters
Vector databases are the infrastructure layer that makes retrieval-augmented generation (RAG) work. When a user asks Claude or GPT-4 a question about a company's internal documents, the system first queries a vector database to find the most relevant passages, then feeds those passages to the language model as context. Without a fast, scalable vector store, RAG systems cannot function in production.
Beyond RAG, vector databases power semantic search (finding results by meaning rather than keywords), recommendation systems (finding similar products or content), duplicate detection, and anomaly detection. Companies like Notion use vector search for their AI features. GitHub Copilot retrieves relevant code using vector similarity. E-commerce platforms use vector databases to power visual search -- upload a photo and find similar products.
The market is growing rapidly. Pinecone raised over $100 million in funding. Zilliz, the company behind Milvus, raised $60 million. In 2026, vector databases are considered essential infrastructure for any organization building AI-powered applications.
Key Takeaway
Vector databases store and search the numerical representations that AI models use to understand meaning, making them the essential infrastructure layer for RAG, semantic search, and recommendation systems.
Part of the AI Weekly Glossary.