ai-stack
The brains and agents stacked on mcp-node: RAG and Container Apps in one capped group.
Overview
Project two in the series, stacked on mcp-node. It adds the service side: a RAG data plane (Postgres with pgvector, plus blob storage for the documents) and a set of agents on Azure Container Apps, all in one capped resource group. It requires a running mcp-node as the control plane.
Architecture
Data plane: Postgres holds the documents, metadata, and vectors in one place; blob storage holds the raw sources; an ingest path embeds them into the vector store. Agent plane: a system Container App (trainer and orchestrator) and a role-based agents app (solution-architect and delivery), both scale-to-zero. Key Vault, a scoped identity, and a budget cap round it out.
Lessons learned
- Postgres with pgvector keeps RAG cheap and in one resource you can actually inspect
- Container Apps scale to zero, but Postgres has an hourly floor, so the cap sits a bit higher
- Chaining project two onto project one keeps each piece small and deployable on its own
Deployment
One-shot deploy into its own resource group on top of a running mcp-node, with a budget cap.