Abstract
Recent advances in RAG (Retrieval-Augmented Generation) focus on optimizing performance at high ingestion costs, such as building knowledge graphs or extracting SQL tables. This work demonstrates that the operations allowed by such knowledge bases can be replicated at zero ingestion costs, without even needing a vector database. Our solution, Zero-Ingestion ScalableRAG, significantly outperforms all baselines (including knowledge graph approaches) in three out of six corpora considered, and only marginally misses maximum performance on the other three, achieving an average accuracy of 7.36% above the next most competitive baseline.
Zero-Ingestion ScalableRAG maintains a workspace of document sets and value sets, enabling on-the-fly aggregative reasoning in scenarios where grouping is necessary based on a primary key corresponding to a subset of the total document set. By capping the number of LLM calls to a constant independent of corpus size, we introduce Limited-Ingestion ScalableRAG, which utilizes a minimal vector database along with automated pattern discovery from a sample of documents to further enhance accuracy at scale.
Our code is available at the following link: ScalableRAG GitHub
Blogger's Review: The Zero-Ingestion ScalableRAG approach proposed in this paper disrupts traditional high-cost knowledge ingestion models by achieving superior performance through efficient document management and real-time reasoning mechanisms. This innovation not only offers new insights for the development of RAG technologies but also presents new possibilities for cost control in practical applications.