LaunchStack treats RAG, OCR, and graphs as swappable ports

What it does

LaunchStack is a TypeScript engine that abstracts the repetitive infrastructure of AI-native apps — document ingestion, OCR, vector search, knowledge graphs, and background jobs — into a framework-agnostic core. You instantiate it with a plain config object and wire up four ports for storage, jobs, credits, and RAG. The included Next.js reference app demonstrates the full stack, but the core library is just an npm package you can drop into any Node 20+ project with Postgres.

The interesting bit

The architecture enforces strict boundaries with ESLint: core knows nothing of the host framework, features depend only on core subpaths, and the host owns routing and auth. It is a ports-and-adapters setup for LLM plumbing, shipping with built-in guardrails like PII filtering and confidence gating.

Key highlights

• Ships as @launchstack/core on npm; runs standalone without the monorepo
• Reads zero environment variables at runtime — config is explicit and portable across Next.js, CLIs, workers, and MCP servers
• Supports multiple OCR providers (Docling, Marker, Azure, Landing.AI, Datalab) and document sources (Notion, Slack, GitHub, Google Workspace exports)
• Includes predictive document analysis across eight types (contract, financial, compliance, etc.)
• Reference app demonstrates role-based auth, pgvector RAG, and marketing pipelines for Reddit, X, LinkedIn, and Bluesky

Caveats

• Several vertical features (MCP server factory, workflow generation, regulatory rule extraction) are on the roadmap but not yet implemented
• Full local stack with OCR enabled needs roughly 1.2 GB RAM; the lite mode is closer to 400 MB
• The reference app requires wiring multiple third-party ports and keys to reach production parity

Verdict

Good for teams tired of rebuilding the same document-to-RAG pipeline for every new AI product. Skip it if you want a fully managed, batteries-included SaaS rather than an engine to wire yourself.