## EmbeddingRWKV: A New Paradigm for RAG A recent study published on arXiv introduces an innovative approach for Retrieval-Augmented Generation (RAG) systems, called State-Centric Retrieval. This new paradigm aims to overcome the inefficiencies of traditional systems, which often perform redundant calculations due to the lack of information sharing between the embedding and reranking stages. The key idea is the use of "states" as a connecting bridge between these two phases. The model at the heart of this approach is EmbeddingRWKV, an RWKV-based model fine-tuned to serve as both an embedding model and a backbone for extracting compact, reusable states. ## Optimized Efficiency and Speed The proposed system also includes a state-based reranker, designed to fully leverage precomputed information. During reranking, the model processes only query tokens, decoupling the inference cost from the document length and achieving a speedup between 5.4x and 44.8x. Furthermore, the researchers found that it is not necessary to retain all intermediate layer states. With a uniform layer selection strategy, the model maintains 98.62% of full-model performance using only 25% of the layers. ## Promising Results Experimental results demonstrate that State-Centric Retrieval achieves high-quality retrieval and reranking results, significantly improving overall system efficiency. The project code is available on GitHub for further investigation and evaluation.