Getting Started
The Modern AI Stack
Last updated March 3, 2026
The Blueprint for building scalable, production-ready AI applications.
When building a modern AI application, you need three foundational layers. We call this the "VSC Stack" (Vercel, Supabase/Neon, Cencori) — or effectively: Frontend, Database, Intelligence.
Framework
Next.js
The React framework. Handles UI, routing, and streaming responses.
Deployment
Vercel
The global edge network. Handles caching, serving, scaling and deployment.
Database
Supabase
The source of truth. Handles auth, structured data (SQL), and relational integrity.
AI
Cencori
The cognitive engine. Handles LLM orchestration, vector memory, security, and agents.
Why separate "Intelligence" from "Database"?
A common mistake is trying to stuff embeddings and AI logic into your primary database (e.g., using pgvector for everything).
While powerful, this mixes concerns:
- Bloat: Vectors are heavy. Mixing them with transactional data slows down backups and migrations.
- Complexity: Managing vector indexes (HNSW/IVFFlat) requires specialized tuning.
- Security: AI prompts need different security rules (PII redaction) than SQL rows.
Cencori acts as the specialized "AI Database", sitting alongside your Postgres "App Database".
The Reference Architecture
Here is how the data flows in a production system:
Capabilities Matrix
| Feature | Provider | Why? |
|---|---|---|
| Rendering | Vercel | Best-in-class React framework & Edge Network. |
| Users & Auth | Supabase | Robust RLS (Row Level Security) & easy Auth helpers. |
| Transactions | Neon | Serverless branching matching Vercel's preview deployments. |
| Memory/RAG | Cencori | Managed Vector Store tailored for high-speed retrieval. |
| Security | Cencori | Dedicated PII Redaction & Prompt Injection firewall. |
| Observability | Cencori | AI-specific traces (what token caused the error?). |
Start Building
Choose your backend integration to get started: