See how much RAM your HNSW index needs, which Rivestack plan fits, and what it would roughly cost to run the same thing yourself. All figures are estimates.
Total rows in your vector table
1536 = OpenAI text-embedding-3-small/large; 768 = Cohere/BGE-base
halfvec / fp16 cuts index RAM roughly in half
1 node has no failover. 2 or 3 nodes add automatic failover. The extra nodes are replicas with the same data, so they buy uptime, not capacity.
Formula (ESTIMATE): vectors × dimensions × 4 bytes (fp32) × 1.3 graph overhead . Actual usage varies with HNSW parameters (m, ef_construction) and data shape.
pgvector pre-installed and HNSW-tuned on dedicated NVMe. Flat pricing, free migration help.
HNSW indexing, NVMe benchmarks, and how Rivestack tunes pgvector for production.
How pgvector HNSW indexes work, when to use IVFFlat, and production tuning parameters.
Why always-on dedicated Postgres beats serverless for latency-sensitive vector workloads.