LLM-Powered RAG Chatbot

The goal was question-answering over a private set of documents, with answers grounded in those documents rather than in whatever a model remembers. A LangChain ingestion pipeline chunks the source files and stores embeddings in Supabase for semantic retrieval, and a local Ollama model writes the answer from the retrieved context. The Next.js chat interface streams responses as they generate and handles the two failure modes that make RAG demos feel broken: retrievals that come back empty, and model calls that time out. Both return a clear message instead of a made-up answer or a hung spinner. This is the project behind my focus on LLM application engineering: building reliable systems around existing models rather than training new ones.

More detailed notes — architecture, decisions, tradeoffs, and outcomes — get published here as the project matures, alongside source code and live deployments.