ScalableRAG Achieves High-Quality RAG with Zero Ingestion Cost

Retrieval-Augmented Generation (RAG) systems typically incur significant "ingestion costs" for building knowledge bases like graphs or SQL tables to optimize performance. This new research presents Zero-Ingestion ScalableRAG, a solution that eliminates these costs while still delivering high-quality RAG. The system operates without even a vector database, instead using a dynamic workspace of document and value sets. This allows for real-time, aggregative reasoning, particularly useful when grouping based on primary keys. Zero-Ingestion ScalableRAG demonstrated superior performance, outperforming all baselines, including knowledge graph approaches, in three out of six evaluated corpora. It achieved an average accuracy 7.36% higher than the next best baseline across all datasets. The paper also introduces Limited-Ingestion ScalableRAG, which uses a minimal vector database and automated pattern discovery to further enhance accuracy at scale, while capping the number of LLM calls independently of corpus size. The code for this innovative approach is publicly available.