Every vector database demo looks fast. The differences show up at scale and in operations.

Start with your constraints: how many vectors, what dimensionality, how fresh must the index be, and what latency can you tolerate. A few hundred thousand vectors behave very differently from a few hundred million.

Recall versus speed is a dial, not a fixed value — approximate indexes trade a little accuracy for a lot of throughput, and the right setting depends on your task. Measure recall on your own data, not on a benchmark.

Then weigh operations: managed service or self-hosted, backup and reindex story, metadata filtering, and cost at your real volume. For many teams, Postgres with a vector extension is enough and keeps everything in one place. Reach for a dedicated store when scale, filtering or latency genuinely demand it.