Do you even need a separate vector database?

Vector similarity search finds semantically related items by comparing numeric representations, and the underlying technique (like Locality Sensitive Hashing) has existed for decades — what's new is the scale driven by embeddings from large language models. Most current vector databases are really key-value stores or search engines like Elasticsearch and OpenSearch with added vector indexing methods such as HNSW or IVF layered on top. These systems combine vector storage, index building, metadata storage, and similarity search with filtering.

The catch is that primary data usually lives in a regular database like PostgreSQL or MongoDB, while vectors get duplicated into a separate vector database. That split creates synchronization headaches, consistency risks, and higher storage costs, plus the operational burden of running yet another distributed system.