How Vector Databases Work

Vector databases store items as high-dimensional vectors and find similar ones using HNSW — a hierarchical navigable small world graph. Instead of scanning everything, they hop through layers of a graph, getting closer with each jump. Click "Search" to watch it find people named like "John."

Layer 2 (sparse) Layer 1 (medium) Layer 0 (dense)

How HNSW Works

HNSW builds multiple layers of a graph. The top layer has few nodes with long-range connections. Each lower layer adds more nodes with shorter-range links. To search, the algorithm starts at the top and greedily hops to whichever neighbor is closest to the query, then drops down a layer and repeats.

Search Progress

Results

Click "Search" to find vectors similar to the name "John"

The Limitation

A vector search finds similar names: John, Jean, Johan, Juan. But each result is just a vector with an attached JSON document. To find John's father, you'd need to read the JSON, extract the father's ID, then do another query. For great-grandfathers? That's three sequential lookups through JSON documents — the vector graph has no concept of relationships.