wtf is controllable-rag-agent?

The controllable-rag-agent project helps you build an AI system that can answer genuinely complicated questions about your own documents, the kind of multi-step questions that simpler AI tools get wrong. Instead of just searching for relevant paragraphs and hoping for the best, it breaks hard questions into smaller pieces, plans how to tackle each one, and verifies its answers along the way. Here is how it works in plain terms. You load your documents (the demo uses the first Harry Potter book as an example). The system creates summaries of each section and stores everything in a searchable format. When you ask a question, it first strips out specific names and details to create a generic version of your question, this prevents the AI from relying on what it already knows and forces it to work only from your documents. Then it builds a step-by-step plan. For a question like "How did the protagonist defeat the villain's assistant?" it would figure out who the protagonist is, who the villain is, who the assistant is, search for relevant interactions, and piece together the answer. After each step, it checks whether the information it found actually came from your documents rather than being invented, and it adjusts its plan as new information comes in. This is useful for anyone who needs an AI to reason carefully over a specific document set, a legal team analyzing case files, a researcher working through a long technical manual, or a product team building an internal knowledge tool where accuracy really matters and hallucinations are unacceptable. The project is built as Jupyter notebooks with a visual demo interface, so you can watch the agent think through each step in real time. It is designed to be controllable rather than a black box: you can see and influence how it plans, retrieves, and verifies. It uses established tools like LangChain and evaluation metrics from Ragas to measure answer quality, and it works with several different AI model providers so you are not locked into one.