Vector databases without the hype
Under a million vectors, the database you already run is probably enough. We measured it.
Vector databases were the new hot topic getting investment and hype. Pinecone, Weaviate, and others were being marketed as the necessary infrastructure for semantic search. We wanted to know if the premise held up to scrutiny and real-world testing.
We built the same index twice: once in a dedicated vector database and once in Postgres with an extension. Same hardware. Same queries. Under a million vectors, which is where most small projects operate.
The dedicated database was slightly faster in latency. The difference in latency was smaller than the variation between regular deploys. Smaller than the difference between weekday and weekend traffic patterns and load.
The operational overhead of running another database was visible and real. One more system to upgrade. One more thing that could fail. One more node to monitor and keep running.
For projects under that threshold, adding Postgres to existing infrastructure beat adding a new specialized system. The latency gain did not justify the operational cost or learning curve.
That comparison became the template for every vector question since. Numbers from your actual data and your actual load, not reference architectures and worst-case benchmarks from marketing materials.