Database

The template uses Prisma 7 with PostgreSQL and the pgvector extension (for RAG embeddings). The Prisma client connects through a driver adapter (@prisma/adapter-pg).

Schema

The schema is split by domain in — Prisma merges every file:

File	Models
`schema.prisma`	generator + datasource (pgvector extension)
`auth.prisma`	`User`, `Session`, `Account`, `Verification`
`billing.prisma`	`Subscription`
`credits.prisma`	`CreditBalance`, `CreditTransaction`
`chat.prisma`	`Conversation`, `Message`
`rag.prisma`	`Document`, `DocumentChunk` (vector column)

Local database

The db service in ships pgvector/pgvector:pg16, so the extension is available out of the box.

bash

docker compose up -d db     # startdocker compose stop db      # stop (keeps data)docker compose down -v      # remove everything, including the volume

docker compose up -d db     # startdocker compose stop db      # stop (keeps data)docker compose down -v      # remove everything, including the volume

Commands

Command	What it does
`npm run db:migrate`	Create & apply a migration in development
`npm run db:deploy`	Apply existing migrations (production / CI)
`npm run db:seed`	Seed the demo account
`npm run db:studio`	Open Prisma Studio (visual browser)
`npm run db:reset`	Drop, re-migrate and re-seed
`npm run db:generate`	Regenerate the Prisma client

Seed

creates a demo account (demo@nova.dev) with a subscription, credits, a conversation and an indexed document — so a fresh clone shows a working product immediately. It is idempotent.

pgvector & RAG

DocumentChunk.embedding is a vector(1536) column. Prisma has no native vector type, so it is declared Unsupported(...) and read/written via raw SQL in the RAG service. An HNSW index (cosine distance) is created in the migration for fast similarity search. The dimension matches OpenAI text-embedding-3-small; change it in the schema and the embedding config if you switch models.