// the find
Deodat-Lawson/LaunchStack
AI-powered StartUp Accelerator Engine built with Next.js, LangChain, PostgreSQL + pgvector. Upload, organize, and chat with documents. Includes predictive missing-document detection, role-based workflows, and page-level insight extraction.
Launchstack is a TypeScript monorepo that puts document ingestion, OCR, pgvector-backed RAG, a knowledge graph layer and background jobs behind ports, with a Next.js app that shows how the pieces fit. It is aimed at developers who want to self-host a document-chat product or reuse the engine in their own app, and who can accept that the packages are not on npm yet.
The layering is enforced, not just documented. ESLint blocks core from reading process.env and features from importing the host app, and CI treats that as a blocking gate. Storage, jobs, credits and RAG are ports, so the host owns the Inngest and S3 choices and the engine stays testable with fakes. The globalThis slot registration is a deliberate answer to Next.js HMR duplicating module state, and the README says so openly. Chat model routing is explicit YAML, and specialized routes such as vision fail closed when no capable model is configured, so images never silently go to a text-only model. The README also documents its own limitations, such as the ZIP adapter skipping JSON exports inside archives, which is more useful than a longer feature list.
The engine is not published. The README says plainly that consumers have to run this repo, so the reusable part is currently a monorepo to fork or vendor, not a dependency to add. Configuration is easy to get wrong on first setup. A bare OPENAI_API_KEY or OPENROUTER_API_KEY does not configure chat, CHAT_BASE_URL defaults to Gemini, and AI_BASE_URL is still translated with a deprecation warning. Each choice has a reason, but a newcomer will hit them. The predictive document analysis is described as covering eight document types, yet the request validator accepts only five. That gap sits on a headline feature, so check the code before relying on it. The scope is also wide for one repo. Marketing pipelines for Reddit, X, LinkedIn and Bluesky, a client prospector, trend search and legal templates share the tree with the engine, and the mcp and connectors packages are empty scaffolding. All of that raises the maintenance cost of the part that matters most.