Dev packages a reusable Next.js RAG starter kit to skip re-wiring pipelines

A reusable starter kit packages the common RAG (retrieval-augmented generation) pipeline: ingesting documents like PDFs or web pages, splitting them into chunks, converting them into vector embeddings, storing them in a vector database, and chatting over them with an LLM. It's built on Next.js, with chat responses handled by the Claude Haiku model streamed via SSE, embeddings generated through Voyage AI, and storage in Pinecone using per-user namespace isolation so different users' data stays separate. Web scraping runs on Cheerio, avoiding the need for a headless browser so it works on serverless platforms.

Rather than a bare demo wrapper, it includes production-grade plumbing accumulated from repeated client work: a specific chunking strategy, similarity thresholds for retrieval, multi-tenant namespace separation, and rate limiting implemented without needing Redis. It deploys to Vercel with just three environment variables.