Vector database

What it is

IBM describes a vector database as one that stores, manages and indexes high-dimensional vector data, so it can run similarity search at scale. Options come in three broad kinds:

• dedicated vector databases;
• extensions to databases you may already run, such as pgvector, an open-source extension that adds vector similarity search to Postgres;
• managed services, such as Cloudflare Vectorize.

IonBin names examples to explain the categories, not to rank them.

How it works

1. Embed each document chunk and store the vector with its text and metadata.
2. At query time, embed the question and ask the database for the nearest neighbors.
3. Pass the matching chunks to the model as context for its answer (RAG).

Benefits

• Fast semantic search over millions of items.
• Metadata filters (date, customer, product) narrow the results.
• Managed options remove infrastructure work, and an extension keeps everything in one existing database.

Limits and risks

• Adds a system to secure, update and pay for.
• Retrieval quality depends on chunking and embeddings, not just the database.
• Switching embedding models means re-embedding every stored item.
• For small collections, simpler approaches often work as well (see do I need a vector database?).