Imagine an AI assistant that can answer any question by searching the vast ocean of online information. That's the promise of Retrieval Augmented Generation (RAG). But what if the AI doesn't know the right way to search? That's where query rewriting comes in, like giving your AI a better search strategy. Existing methods for training AI to rewrite queries often rely on manual labeling or pre-defined rewards, which can be costly and inflexible. A new research paper introduces RaFe, a clever framework that uses feedback from a 'reranker' to train query rewriting models. A reranker is like a judge that evaluates how well different search queries retrieve relevant documents. RaFe uses this feedback to guide the AI towards generating better search queries, without needing labeled data. This innovative approach has shown promising results in open-domain question answering across both English and Chinese datasets. By using a publicly available reranker, RaFe demonstrates its potential for broad application. The research also explores how many rewrites are optimal for balancing performance and efficiency. It turns out that just a few well-crafted rewrites can significantly boost the AI's ability to find the right information. This research opens exciting new avenues for improving how AI interacts with information, paving the way for more accurate and efficient retrieval systems.