Choosing a Vector Database: What the Benchmarks Don't Tell You

Every vector database publishes benchmarks showing strong recall and low latency, and at the scale most business applications actually operate at, those numbers converge close enough to be a rounding error in the decision. The differentiators that actually matter show up in production, not in the benchmark suite.

• Metadata filtering performance at query time — most real queries aren't pure similarity search, they're similarity plus filters.
• Hybrid search (keyword plus vector) support out of the box versus bolted on separately.
• Operational story: managed service versus self-hosted, and what a re-index of ten million vectors actually costs in time and money.
• Multi-tenancy support if you're serving many customers from one index — not every vector database isolates this cleanly.

I've shipped systems on Pinecone, pgvector, and Elasticsearch's vector support, and the deciding factor was never raw recall — it was which one fit the team's existing operational skills and the filtering complexity the actual queries needed.

Pick the vector database that fits your filtering needs and your ops team's existing skills. The recall numbers will be fine either way.