A minimal Express + TypeScript backend that talks to a VelesDB server through the @wiscale/velesdb-sdk Node.js client.

Three endpoints:

cd examples/node-express-rag
docker compose up --build

This starts:

• velesdb-server (built from the repo root Dockerfile) on :8080 with persistent volume velesdb-data
• node-express-rag (this app) on :3000, configured via VELESDB_URL=http://velesdb:8080

The app waits for the VelesDB healthcheck to go green before accepting traffic.

First build is slow: docker compose up --build compiles the full Rust workspace (~5–15 minutes cold). Subsequent runs use the cached image. If you already built velesdb/velesdb:latest from the repo root, change the velesdb service in docker-compose.yml to image: velesdb/velesdb:latest (drop the build: block) to skip the rebuild. The named volume velesdb-data survives docker compose down but is wiped by docker compose down -v.

# Ingest a few documents
curl -X POST http://localhost:3000/ingest \
  -H "content-type: application/json" \
  -d '{"id":"1","text":"VelesDB is a single-binary embedded vector + graph database in Rust."}'

curl -X POST http://localhost:3000/ingest \
  -H "content-type: application/json" \
  -d '{"id":"2","text":"Express is a minimal Node.js web framework for building HTTP APIs."}'

curl -X POST http://localhost:3000/ingest \
  -H "content-type: application/json" \
  -d '{"id":"3","text":"Cosine similarity ranks documents by the angle between their embedding vectors."}'

# Search
curl -X POST http://localhost:3000/search \
  -H "content-type: application/json" \
  -d '{"query":"how does similarity search work","k":3}'

# Health
curl http://localhost:3000/health

If you have a velesdb-server process running on the host (cargo run --bin velesdb-server):

cd examples/node-express-rag
cp .env.example .env       # adjust PORT / VELESDB_URL if needed
npm install
npm run dev                # tsx watch mode

src/embed.ts ships a deterministic 384-D hashing-trick projection (the same trick used by examples/react-wasm-search/). It is intentionally a placeholder so the example is self-contained — no model download, no external API key — but it only captures lexical overlap, not semantics.

Production-ready alternatives:

• In-process: @xenova/transformers (ONNX runtime)
• Remote: any embeddings API that returns a fixed-size vector

Swap the embed() function and bump EMBEDDING_DIM to your model's output size; the rest of the pipeline is unchanged.