Managed pgvector hosting
on dedicated NVMe.
Vector search in the Postgres you know. pgvector, local NVMe, backups, and monitoring, ready for your app.
From $29/month · EU Central, US East, and Singapore
~1,185
QPS
4 clients
3.2
ms p50
4 clients
0.93
recall@10
ef_search=80
See the benchmarks below for full results and methodology.
Local NVMe
dedicated per node, no IOPS cap
2-5 ms
p50 vector search, recall@10 0.90+
$29/mo
starting, fixed pricing
EU
Hosted in Germany · GDPR-ready
01 — Why managed
Why managed pgvector
We handle index tuning, backups, and database operations as your workload grows.
NVMe-grade vector search
Local NVMe keeps random reads fast as your vector index grows beyond memory.
NVMe vs cloud SSD benchmarks →pgvector pre-installed and tuned
pgvector 0.8.x is installed, with HNSW defaults and tuning guidance for your workload.
pgvector quickstart →Backups, HA, and metrics, included
Backups, recovery, and monitoring are included. Add at least two nodes for automatic failover.
Why managed Postgres for AI →02 — Providers
Best managed pgvector hosting for production AI apps
The right provider is not just the one that can run CREATE EXTENSION vector . Compare memory, storage, upgrades, and support against your workload.
Best managed pgvector provider for production RAG
Compare storage, memory, index tuning, and support for your workload.
Compare providers →
Best pgvector hosting for European data residency
Choose EU Central, US East, or Singapore. Review database, backup, and support locations for your requirements.
Read the hosting guide →
Best Supabase or Neon alternative for pgvector
Compare the database features you use, the capacity you need, and the total monthly cost.
See the comparison →
03 — Compare
Rivestack vs Exoscale vs Elestio vs Supabase
Compare the defaults that matter for vector workloads.
| Rivestack | Exoscale | Elestio | Supabase | |
|---|---|---|---|---|
| Best fit | Postgres AI workloads | European cloud buyers | Marketplace-style hosting | Full app platform |
| Starting price | $29/mo · Solo: 7-day free trial for new customers (card required) | Hourly DBaaS | $14/mo+ | $25 + compute add-ons |
| Storage default | Dedicated local NVMe | Managed DBaaS storage | Provider-dependent VM disk | Cloud block storage |
| pgvector support | 0.8.x line | Extension included | Extension service | Extension included |
| HNSW tuning | Workload tuned · ef_search guidance | General DB tuning | Managed service defaults | Extension defaults |
| Vector workload review | Included | Support plan | Support plan | Community / paid |
| Pricing model | Fixed per node | Hourly usage | Monthly service | Plan + usage/add-ons |
| Migration path | pg_dump, restore, or logical replication help | Standard PostgreSQL tools | DB migration service | Import/export tooling |
Provider pages and prices change; use this as a buying checklist and verify final plan details before committing. Vector benchmarks measured per tier on 250k to 1M × 1536-dim, HNSW m=16 / ef_construction=64 on AMD cpx nodes (Starter ~1,600 QPS at recall 0.93; Scale serves 1M hot at ~3,600 QPS). Full measured, reproducible tables →
04 — Benchmarks
Measured pgvector benchmarks, per tier
1536-dim vectors, measured from a same-region app over PgBouncer + TLS. Every figure carries its recall and client count.
| Plan | Throughput | p50 | recall@10 | Clients | Served hot |
|---|---|---|---|---|---|
| Solo1 vCPU · 2 GB | ~983 QPS | 3.8 ms | 0.93 | 4 | ~250K |
| Growth4 vCPU · 8 GB | ~2,950 QPS | 5.1 ms | 0.94 | 16 | ~600K |
| Scale8 vCPU · 16 GB | ~4,465 QPS | 3.4 ms | 0.95 | 16 | ~1M |
Per-tier detail, test configuration, and methodology
solo
~983 QPSThroughput
250k × 1536 at recall@10 0.93 (ef_search=80), 4 clients, p50 3.8 ms (a single client is ~1.3 ms). Every operating point is at /pgvector-benchmarks-measured
1.3 msp50 Latency
Single client at recall@10 0.93; 3.8 ms at 4 clients. Same-region over PgBouncer + TLS
~250KHonest capacity
1 vCPU / 2 GB. The 250k HNSW index builds in ~15 min and stays hot; larger indexes are RAM-limited, so use Starter or above
growth
~2,950 QPSThroughput
250k × 1536 at recall@10 0.94 (ef_search=80), 16 clients, p50 5.1 ms, and ~1,990 QPS at 4 clients, p50 1.9 ms
~2,570 QPSHigh recall
250k at recall@10 0.95 (ef=120), 16 clients, p50 5.8 ms, so throughput stays high even at 0.95 recall
~600KHonest capacity
4 vCPU / 8 GB. The 250k index builds in ~2 min; a full 1M index needs Scale
scale
~4,465 QPSThroughput
250k × 1536 at recall@10 0.95 (ef_search=80), 16 clients, p50 3.4 ms, up to ~5,800 QPS at lower recall
~3,600 QPS1M served hot
1M × 1536 at recall@10 0.74 (ef=80), 16 clients, p50 4.2 ms on an 8-vCPU / 16 GB node; the ~6 GB index builds in ~9 min and stays hot in RAM
~1MHonest capacity
250k builds in ~80s. 1M × 1536 (~6 GB) stays hot on the 16 GB node; smaller nodes go disk-bound
Solo (2 GB) serves ~250K hot, Growth (8 GB) ~600K, and Scale builds and serves 1M × 1536 hot at ~3,600 QPS / p50 ~2 ms (recall 0.75). QPS and p50 trade off across the concurrency curve. OpenAI text-embedding-3-small (1536 dimensions), same-region client. Full methodology, the ef_search/recall sweep, and reproduce-it steps in the pgvector NVMe vs cloud SSD benchmarks, with every operating point at /pgvector-benchmarks-measured.
05 — How it works
From zero to an HNSW query in three steps
From "we should try pgvector" to a working HNSW query in three steps, no extension wrangling.
Create a database in 30 seconds
Choose your region and node size. We provision and configure the database.
pgvector and pgvectorscale are enabled
pgvector and pgvectorscale arrive enabled, with HNSW, IVFFlat, and StreamingDiskANN support.
Choose an index and query
Create an index and query it with your existing pgvector client and SQL.
CREATE TABLE docs (
id BIGSERIAL PRIMARY KEY,
text TEXT,
emb vector(1536)
);
CREATE INDEX ON docs
USING hnsw (emb vector_cosine_ops);
SELECT id, text, emb <=> $1 AS dist
FROM docs
ORDER BY emb <=> $1
LIMIT 10;Need a deeper walkthrough? See the full pgvector getting-started guide and the pgvector vs Pinecone comparison.
06 — Migrate
Migrate to managed pgvector without rewriting your app
Keep your pgvector queries. We help move the data and plan the cutover.
Supabase, Neon, RDS, Aiven, or Exoscale
Move tables, indexes, and roles with standard PostgreSQL tools. Plan and validate the cutover first.
Pinecone or another vector database
Export embeddings and metadata, then map them to vector columns and SQL filters.
Self-hosted pgvector
Choose a migration method around your downtime needs. We take over backups, monitoring, and operations.
07 — Pricing
Fixed monthly pricing, per node
One price per node. No fees for queries, egress, or vector count.
Prices shown are EU Central, our lowest region. US-East and Singapore are higher. Full pricing.
Your own single-tenant VM for small production apps. Fully backed up, 7-day free trial for new customers.
Start free trial- · 7-day free trial (card required)
- · 1 vCPU · 2 GB RAM
- · 30 GB NVMe storage
- · ~250K vectors (1536d)
- · pgvector tuned (HNSW)
- · Daily backups + 7d PITR
- · Monitoring dashboard
- · Never deleted
Production-ready dedicated PostgreSQL with NVMe storage, automated backups, and monitoring.
Start Starter- · 2 vCPU · 4 GB RAM per node
- · 55 GB NVMe storage
- · ~300K vectors (1536d)
- · pgvector tuned (HNSW)
- · HA-ready (Patroni failover)
- · Daily backups + 14d PITR
- · Monitoring dashboard
- · Terraform provider
More compute, more storage. Same HA support, add nodes for automatic failover.
Start Growth- · 4 vCPU · 8 GB RAM per node
- · 135 GB NVMe storage
- · ~600K vectors (1536d)
- · pgvector tuned (HNSW)
- · HA-ready (Patroni failover)
- · Daily backups + 14d PITR
- · Priority support
- · Custom PostgreSQL config
High-performance dedicated PostgreSQL for demanding vector workloads.
Start Scale- · 8 vCPU · 16 GB RAM per node
- · 295 GB NVMe storage
- · ~1M vectors (1536d)
- · pgvector tuned (HNSW)
- · HA-ready (Patroni failover)
- · Daily backups + 14d PITR
- · Priority support
- · Custom PostgreSQL config
08 — FAQ
Managed pgvector FAQ
What teams ask before moving a real pgvector workload.
PostgreSQL with pgvector installed and operated for you. You use standard vector types, indexes, and SQL; we handle backups, monitoring, and database operations.
Yes. Plan for index memory, storage latency, recall targets, and filters. Benchmark your own data before choosing a node.
Size for index memory and latency, not row count alone. For 1536 dimensions, our measured capacity ranges from about 250K vectors on Solo to 1M on Scale. Larger indexes can spill to disk; see the benchmark tables for conditions.
HNSW uses a graph to find similar vectors. The m and ef_construction settings shape the index; hnsw.ef_search trades query speed for recall. Memory and storage latency both matter.
Often. pgvector keeps vectors, filters, and relational data in one database. A dedicated vector service may suit very large datasets or workloads with no relational context.
Compare storage latency, index memory, supported versions, backups, availability, pricing, and migration help against your actual workload.
Performance depends on node size, recall, concurrency, and index memory. The benchmark table above includes these conditions; the measured benchmarks page provides full runs and reproduction steps.
Yes. PostgreSQL sources use standard migration tools. Pinecone requires exporting embeddings and metadata into PostgreSQL tables. We help plan the mapping and cutover.
We support pgvector 0.8.x. Updates follow compatibility testing and planned maintenance windows.
Yes. Rivestack supports PostgreSQL 16, 17, and 18, with 17 as the current default. pgvector and pgvectorscale are included.
Both. HNSW is our default. IVFFlat can reduce build time and memory for suitable datasets. You can use both in one database.
09 — Guides
Managed pgvector buying guides
More detailed comparisons for teams choosing a pgvector provider.
Best managed pgvector providers
How to compare Rivestack, Exoscale, Elestio, Supabase, Neon, Aiven, AWS, and Google Cloud for production vector search.
Read guide →
pgvector hosting guide
A practical checklist for choosing hosted pgvector: storage, memory, HNSW tuning, backups, HA, and migration support.
Read guide →
Supabase pgvector alternative
When Supabase is perfect, and when a focused managed PostgreSQL service is a better fit for vector-heavy workloads.
Read guide →