Large language models (LLMs) have revolutionized various natural language processing tasks. However, their limited parametric memory makes them susceptible to generating inaccurate or non-factual content. Retrieval-augmented generation (RAG) addresses this limitation by integrating information from external sources. Yet, these systems often encounter challenges with irrelevant information in retrieved documents. Addressing this, researchers have developed DSLR, an unsupervised document refinement framework designed to enhance RAG. DSLR employs a three-step process: decomposition, re-ranking, and reconstruction. It breaks down retrieved documents into sentences, assesses their relevance to the query, and reassembles them into coherent passages, eliminating irrelevant sentences along the way. This approach improves the precision of information delivered to the LLM, thus enhancing the quality of generated text. Extensive evaluation across a range of question-answering datasets demonstrates DSLR's significant performance improvement in RAG systems. It streamlines information delivery to the LLM, leading to more accurate and efficient text generation. Notably, DSLR achieves this without the need for additional training, making it a highly adaptable solution for various applications.