If your product touches LLMs, you have probably encountered vector databases. They are the storage layer for embeddings, the dense numerical representations that let computers reason about similarity instead of exact matches.
What They Actually Do
A vector database indexes high-dimensional vectors so you can find the ones most similar to a query, fast. That is the entire pitch.
When You Need One
- Retrieval-augmented generation over a corpus
- Recommendation systems
- Semantic search across documents or code
When You Do Not
If your corpus is small and your queries are exact, a regular database with a text index is faster and simpler. Many teams reach for a vector DB before they need one.
Choosing One
Look at recall, latency, cost per million vectors, and how the index handles updates. The fast-growing tier of options means yesterday's benchmarks are stale.
Conclusion
Vector databases are a powerful new primitive. Adopt them deliberately, not by default.