Arinagi-web/Offline-RAG ? reverse-engineered prompt

Build me a local first document question answering app that works fully offline after the models are downloaded.

I want to be able to upload PDFs, DOCX files, and images, then have the app extract the text, use OCR when needed, add a short caption for images, split everything into searchable chunks, and let me ask questions against my own files. It should retrieve the most relevant chunks, answer with sources, and support both normal and streaming responses. Please include a simple web UI so I can upload files, search my workspace, and see basic usage stats like recent chats, latency, and cache hits.

Keep it privacy friendly, no remote model calls during runtime, and make it run on either CPU or GPU if available. If anything is unclear, look up current docs online if you need to, but keep the final result practical and easy to run.