Deploy and Host Vespa on Railway

Vespa is an open-source search and vector database engine that serves text (BM25), vector (HNSW nearest-neighbor) and hybrid queries from one system. This template runs a single-node Vespa from a digest-pinned image with a sample application already deployed, so the document API and the search API answer as soon as the service is healthy.

About Hosting Vespa

Vespa expects to be its own config server and container, and it runs as a non-root user. The start command handles what Railway needs:

Writable state. Railway mounts volumes root-owned. The service starts as root, fixes ownership of /opt/vespa/var, then drops to the vespa user. Config, indexes and documents persist across redeploys.
Sample application. On first boot a small application is deployed: schema doc with title and text (BM25) and a 384-dimension embedding (HNSW, angular), plus rank profiles bm25 and hybrid. If an application is already on the volume, it is left untouched on later boots.
Single-node hostname. The node name is pinned so the config server and container find each other after every restart.
Public API on 8080. The document and query API is served on Railway's injected port. The config server (19071) stays internal.

To use your own schema, deploy your application package to the config server over the private network.

Why Deploy

Hybrid keyword plus vector search in one engine, with no separate vector store
Ready to feed: no application package to build before the first document
Data survives redeploys on the attached volume
Image pinned by digest, so a deploy never changes under you

Common Use Cases

Retrieval for RAG: store chunks with embeddings and query by vector, keywords or both
Semantic site or product search
Recommendation and similarity lookups
A self-hosted alternative to managed vector databases

Dependencies for Vespa Hosting

A persistent volume, mounted at /opt/vespa/var (created by the template)

Deployment Dependencies

Vespa: https://github.com/vespa-engine/vespa
Documentation: https://docs.vespa.ai
Image: https://hub.docker.com/r/vespaengine/vespa

Implementation Details

Feed a document:

curl -X POST "$VESPA_URL/document/v1/default/doc/docid/1" \
  -H 'Content-Type: application/json' \
  -d '{"fields":{"title":"hello","text":"vespa on railway","embedding":{"values":[/* 384 floats */]}}}'

Keyword query: curl "$VESPA_URL/search/?yql=select%20*%20from%20doc%20where%20userQuery()&query=railway&ranking=bm25"

The data plane has no authentication. Before using it with real data, put it behind your own gateway or reach it only over Railway's private network. Vespa stops accepting writes when its disk is more than 80% full, so size the volume with headroom.