Zero-Ingestion RAG Outperforms Knowledge Graph Approaches
A new paper on arXiv (2607.25135) introduces ScalableRAG, a retrieval-augmented generation system that achieves high-quality results without any ingestion cost. The method, called Zero-Ingestion ScalableRAG, eliminates the need for vector databases, knowledge graphs, or SQL tables. It maintains a workspace of document sets and value sets that can be written to and read from, enabling on-the-fly aggregative reasoning when grouping is required on a primary key in one-to-one correspondence with a subset of the total document set. In evaluations across six corpora, ScalableRAG outperformed all baselines, including knowledge graph approaches, on three datasets and achieved near-maximum performance on the other three. Its average accuracy across all six datasets was 7.36% above the next most competitive baseline.
Key facts
- ScalableRAG achieves zero ingestion cost.
- No vector database, knowledge graph, or SQL tables needed.
- Uses a workspace of document sets and value sets.
- Outperforms all baselines on three of six corpora.
- Near-maximum performance on the remaining three corpora.
- Average accuracy 7.36% above next best baseline.
- Enables on-the-fly aggregative reasoning.
- Published on arXiv with ID 2607.25135.
Entities
Institutions
- arXiv