Neo4j

Graph database (nodes/relationships/Cypher), hosted on Neo4j Aura. Backs the Graph RAG retrieval pipeline. See the db-graph skill for Cypher patterns.

Why was it chosen?

• 5+ level traversals stay linear where recursive SQL goes exponential — the RAG pipeline walks the entity graph deeply.

See vendors.md for pricing, free tier limits, and provider alternatives.

Known limitations

• No native horizontal sharding — vertical scaling only; single writer.
• Relationship-type limit: 65K maximum.
• Aura free tier: one instance.

Aura free tier: auto-pause & resume

A free Aura instance pauses after a period of inactivity. The first query after a pause pays a resume cost on top of normal query latency:

Live at /showcases/db/graph/connection — click Re-test right after a cold start to watch the same instance settle into the Warm tier.

Shared instance, tenant isolation

Aura free = one database, so a single instance holds two graphs at once:

• the public "Tech Stack" demo graph — labels Layer, Technology, Concept, Showcase
• per-user RAG nodes — :Entity / :Chunk, owned via an ownerId property

A public, unauthenticated showcase visitor must never read another user's RAG data. Two guards enforce that:

• Label-scoping. Every public showcase query (src/lib/server/graph/showcase/queries.ts) is constrained to the four demo labels — no CALL db.labels(), no unscoped MATCH (n). A leaked :Entity id returns null.
• Admin gating. The arbitrary-Cypher REPL and the reseed action can't be label-constrained, so both are gated behind isAdmin.

RAG retrieval paths (graph/rag/queries.ts + the rawrag tiers) were already owner-scoped by ownerId — no change there.

Don't copy this naively. An emulating project must either run the demo graph on a separate instance or replicate the label-scoping. A shared instance with unscoped public queries leaks tenant data.

Related

• postgres.md - Relational data
• drizzle.md - PostgreSQL ORM
• ../ai/ai-sdk.md - Graph RAG integration