Fine-tuning large language models (LLMs) is highly resource-intensive, making scalable alternatives like Retrieval-Augmented Generation (RAG) increasingly attractive. However, naive RAG implementations often suffer from limited retrieval precision, long processing times, and high resource costs. We propose KeyKnowledgeRAG ( \(K^2RAG\) ), a scalable framework that integrates dense and sparse retrieval, knowledge graphs, and summarization into a unified pipeline using a divide-and-conquer strategy. \(K^2RAG\) also includes a pre-processing step that reduces training load and enhances chunk relevance. Evaluated on the MultiHopRAG benchmark, our method achieves a mean answer similarity of 0.57 and top-quartile similarity of 0.82, while reducing training time by 93%, execution time by 40%, and VRAM usage by 66%. These results demonstrate that \(K^2RAG\) offers a scalable and efficient solution for developing lightweight, robust question-answering systems based on internal documents while offering excellent answer accuracy and quality.