Postgres 18 + pgvector Vector Database

Deploys PostgreSQL 18 with pgvector 0.8 on Railway's own Postgres image, with PgBouncer connection pooling and daily verified pg_dump backups to a Railway Bucket. A drop-in Postgres service for storing and querying embeddings with SQL.

What you get

  • PostgreSQL 18 with pgvector and pg_stat_statements, TLS, a volume, and a TCP proxy for external clients
  • The vector extension created on first boot; memory and parallel workers sized from the container limits at every start
  • PgBouncer 1.26.1 in transaction mode on the private network, with a pooled DATABASE_POOL_URL
  • A daily cron backup at 03:15 UTC of roles and every database, verified, kept 14 days with at least 3 runs retained
  • A Railway Bucket for the backups; connection variables follow Railway's Postgres layout

Good for

  • Retrieval-augmented generation: document chunks, metadata and embeddings in one database
  • Semantic search over products, tickets, docs or chat history with HNSW or IVFFlat indexes
  • A vector store for LangChain, LlamaIndex, Mem0 or your own app without a separate vector database
  • A general-purpose Postgres 18 for an app that will add embeddings later

Services in this template

  • Postgres · image ghcr.io/railwayapp-templates/postgres-ssl:18
  • Postgres Backup · from baranberkay96/pgvector-railway
  • PgBouncer · image edoburu/pgbouncer:v1.26.1-p0

Questions people ask

How do I connect?

Use DATABASE_URL for direct private connections (migrations, CREATE INDEX, LISTEN/NOTIFY), DATABASE_POOL_URL for application queries through PgBouncer, and DATABASE_PUBLIC_URL from outside Railway with ?sslmode=require. The certificate is self-signed. Create a dedicated role for your app rather than using the postgres superuser, and remove the TCP proxy in Settings, Networking if you want a private-only database.

What drives the cost?

Postgres memory and its volume, which grows with data and indexes. HNSW index builds are fastest when the graph fits in maintenance_work_mem, so large indexes call for more RAM. Backups add Bucket storage, and uploads count as service egress. The template notes estimate roughly $1.50 a month idle and about $16 for a few GB with daily backups; these are estimates.

How do backups and restores work?

Dumps land in the Bucket under postgres/timestamp/ with a MANIFEST.txt. You can trigger one with railway cron run --service "Postgres Backup". To restore, download the database dump from the Bucket and run pg_restore with a PostgreSQL 18 client.

How do I upgrade?

Postgres follows Railway's postgres-ssl:18 tag, with minor and security updates on redeploy; after a pgvector update run ALTER EXTENSION vector UPDATE. PgBouncer is pinned to 1.26.1, so change the tag to upgrade. Use Railway's built-in pooling or this PgBouncer, not both.

Any licence notes?

PostgreSQL and pgvector use the PostgreSQL Licence and PgBouncer uses ISC. This is a community template, not affiliated with the PostgreSQL project, pgvector or Railway.

Full template README

Deploy and Host pgvector with Railway

pgvector adds vector similarity search to PostgreSQL, so embeddings live next to the rest of your data and are queried with SQL. This community template runs PostgreSQL 18 with pgvector 0.8 on Railway's own Postgres image, with PgBouncer connection pooling and daily backups to a Railway Bucket.

About Hosting pgvector

A production pgvector database needs more than the extension. Index builds are fast only when the HNSW graph fits in maintenance_work_mem, parallel builds need enough shared memory, many app instances need a connection pooler, and the data needs off-volume backups. This template keeps Railway's Postgres service layout and variables, so it is a drop-in replacement, and adds those pieces: the vector extension is created on first boot, memory and parallel workers are sized from the container's limits at every start, /dev/shm is raised for parallel index builds, PgBouncer pools connections in transaction mode, and a cron job writes verified pg_dump backups with retention.

Common Use Cases

  • Retrieval-augmented generation (RAG): store document chunks, metadata and embeddings in one database and retrieve with SQL filters
  • Semantic search over products, tickets, docs or chat history with HNSW or IVFFlat indexes
  • Recommendations and deduplication with cosine, inner product or L2 distance
  • A vector store for frameworks such as LangChain, LlamaIndex, Mem0 or your own app, without a separate vector database
  • A general-purpose Postgres 18 for an app that will add embeddings later

Dependencies for pgvector Hosting

  • PostgreSQL 18 with pgvector (included: ghcr.io/railwayapp-templates/postgres-ssl:18)
  • PgBouncer for connection pooling (included: edoburu/pgbouncer 1.26.1)
  • A Railway Bucket for backups (included, created by the template)

Deployment Dependencies

Implementation Details

Service Image Purpose
Postgres ghcr.io/railwayapp-templates/postgres-ssl:18 PostgreSQL 18 + pgvector + pg_stat_statements, TLS, volume at /var/lib/postgresql/data, TCP proxy for external clients
PgBouncer edoburu/pgbouncer:v1.26.1-p0 Transaction pooling on the private network (port 6432), SCRAM auth, prepared statements
Postgres Backup built from this repo (services/backup) Daily cron (03:15 UTC): roles + every database with pg_dump, verified, 14-day retention (minimum 3 runs kept)
Bucket Railway Bucket Private S3 storage for the backups

Connection strings (on the Postgres service, reference them as ${{Postgres.NAME}}):

  • DATABASE_POOL_URL: pooled, for application queries from other Railway services
  • DATABASE_URL: direct over the private network, for migrations, CREATE INDEX, LISTEN/NOTIFY, advisory locks and session settings
  • DATABASE_PUBLIC_URL: from outside Railway through the TCP proxy; add ?sslmode=require (the certificate is self-signed). Remove the TCP proxy in Settings → Networking if you want a private-only database.

First steps

  1. Copy DATABASE_PUBLIC_URL from the Postgres service and connect: psql "{DATABASE_PUBLIC_URL}?sslmode=require".
  2. SELECT extversion FROM pg_extension WHERE extname = 'vector'; confirms pgvector is ready.
  3. Create a table and an index: CREATE TABLE items (id bigserial PRIMARY KEY, embedding vector(1536)); then CREATE INDEX ON items USING hnsw (embedding vector_cosine_ops);
  4. Query with pgvector's cosine-distance operator (see "Querying" in the pgvector README): order by the distance between embedding and your query vector, with LIMIT 10.
  5. Create a dedicated role for your app instead of using the postgres superuser.

Tuning: the start command writes pgvector-tuning.conf at every boot from the container's memory and CPU limits (auto budget capped at 8 GB) and logs the values. Override any setting with ALTER SYSTEM, or set PG_MEMORY_LIMIT_MB / PG_CPU_LIMIT after resizing the service. For a large index build, run SET maintenance_work_mem = '2GB'; on DATABASE_URL; keep it below 3/4 of /dev/shm (RAILWAY_SHM_SIZE_BYTES, 1 GiB) or also SET max_parallel_maintenance_workers = 0;. Through the pool, use SET LOCAL hnsw.ef_search = 100 inside a transaction; Prisma needs ?pgbouncer=true.

Backups and restore: dumps land in the Bucket under postgres/{timestamp}/ with a MANIFEST.txt. Trigger one now with railway cron run --service "Postgres Backup". To restore, download {database}.dump from the Bucket's Files tab and run pg_restore --clean --if-exists --no-owner --dbname "{DATABASE_PUBLIC_URL}?sslmode=require" railway.dump with a PostgreSQL 18 client.

Scaling: give Postgres more RAM and CPU in its settings, then set PG_MEMORY_LIMIT_MB to match so the caches grow with it. Raise DEFAULT_POOL_SIZE on PgBouncer for more concurrent queries (Postgres allows 500 connections). Hobby volumes start at 5 GB; Pro volumes at 50 GB and can be resized. Use Railway's built-in pooling or this PgBouncer, not both.

Versions and upgrades: Postgres follows Railway's postgres-ssl:18 tag (minor and security updates on redeploy; after a pgvector update run ALTER EXTENSION vector UPDATE;). PgBouncer is pinned to 1.26.1; change the tag to upgrade. For Railway's point-in-time recovery or HA, clear the Postgres start command first; the last tuning file stays in place.

Cost: roughly $1.50 per month idle and about $16 per month for a few GB of data with daily backups at Railway's usage prices.

Why Deploy pgvector on Railway?

Railway runs the database, the pooler and the backup job together on a private network with persistent volumes and managed buckets, so you get a vector-ready Postgres in a few minutes without operating servers. It uses Railway's own Postgres image, works with the Data panel and existing ${{Postgres.*}} references, and grows by raising the service limits and the volume size.

Guides for this template

View Postgres 18 + pgvector Vector Database on railway.com

Similar templates

TemplateCategoryDeploymentsDeploys
PostgreSQL
PostgreSQL database service
Storage394,257
MySQL
Deploy a MySQL database service
Storage269,034
MongoDB
MongoDB database service
Storage40,707
Redis
Deploy and Host Redis with Railway
Storage25,288
Redis (Bitnami)
Redis key-value data store
Storage15,087
pgvector
Open-source vector similarity search for Postgres
Storage2,077