ggozad/haiku.rag — reverse-engineered prompt

Build me a local document knowledge assistant in Python. I want to drop in PDFs, documents, and maybe web links, have the app index them, then let me search and ask questions in normal language.

It should answer with citations, including page numbers and section names when possible, so I can trust where the answer came from. It should support chat with memory, not just one question at a time. If a document has charts or figures, I want the system to keep those and use them when answering, including finding images from a text search.

Please include a command line tool, a simple chat interface, and a basic web app for conversations. I also want it to watch a folder and automatically add new files. Keep it local first with an embedded database, but make it configurable for common AI providers for embeddings and answers. If useful, expose the document search and question answering as tools for Claude Desktop through MCP.