What Is a Vector Database? Storing and Searching Embeddings Explained

A vector database is a store built to index and search high-dimensional embedding vectors by similarity, powering semantic search and retrieval-augmented generation.

Modern AI represents text, images, and other data as embeddings - long lists of numbers that capture meaning. To find similar items, you search these vectors by closeness, not by exact match. A vector database is purpose-built for exactly that: fast similarity search over millions of embeddings.

What a vector database is

A vector database stores embedding vectors and lets you query for the nearest ones to a given vector. Instead of matching keywords, it finds items whose meaning is closest. Examples include Pinecone, Weaviate, Qdrant, Milvus, and pgvector for PostgreSQL.

How similarity search works

Comparing a query against every stored vector is too slow at scale, so vector databases use approximate nearest neighbor (ANN) indexes like HNSW. These trade a tiny amount of accuracy for enormous speed, returning the closest vectors in milliseconds across millions of entries.

What they power

Vector databases underpin semantic search, recommendation, deduplication, and retrieval-augmented generation (RAG), where an LLM retrieves relevant documents by embedding similarity before answering. They are a core piece of modern AI application stacks.

Vector databases in CI

Code that embeds data and queries the database is tested in CI against a local or ephemeral vector store seeded with sample vectors.

Testing vector search in CI

steps:
  - run: docker compose up -d qdrant   # ephemeral vector DB
  - run: pytest tests/retrieval        # seed vectors, assert top-k

Latchkey note

Generating embeddings to seed a test index can be GPU-accelerated and slow. On Latchkey, run embedding-heavy tests on GPU or larger runners, cache the embedding model and sample vectors between runs, and auto-retry the transient pulls of the vector-store image or model weights.

Key takeaways

A vector database indexes and searches high-dimensional embeddings by similarity rather than exact match.
Approximate nearest neighbor indexes like HNSW make similarity search fast at scale.
Vector databases power semantic search and RAG, and are tested in CI against an ephemeral seeded store.