We didn’t include pgvector because we focused on managed services to keep things comparable — TopK is managed/serverless, so the fair match would be a managed Postgres. And pgvector just doesn’t really scale to the kinds of workloads we ran here.