ScalableRAG authors propose exact corpus operations without costly ingestion

ScalableRAG authors propose exact corpus operations without costly ingestion

Authors Hilaf Hasson, Aditya Chakravarty, Jayant Thomas, and Krishna Gogineni introduce ScalableRAG, a system that lets an agent build and combine named document and value sets at query time. They argue this closes a gap in retrieval-augmented generation (RAG): lightweight retrievers can find evidence but do not expose the corpus-wide operations needed for counts, intersections, and aggregations, while systems that support such operations require expensive ingestion. The authors report that Zero-Ingestion ScalableRAG needs no preprocessing or vector index and matches the leading baseline on four conventional question-answering benchmarks. On the aggregative Hotels-50 benchmark, they report a score of 71.74, while removing the workspace lowers the score to 38.4. A second version, Limited-Ingestion ScalableRAG, adds a lightweight index; the authors report results above a judge score of 80 at five million documents, with 13 to 16 seconds per question and under 20 minutes of ingestion. The paper’s central argument is that a query-time workspace can combine exact corpus operations with low ingestion cost and scale to large collections. These are reported benchmark results, not evidence here of performance across other workloads. How well would the approach perform on real-world collections and questions beyond the reported benchmarks?

Publication: bbbbae51-e1f7-4b10-8fa9-99d50e8d74f0

attribution — pass: The reading correctly names the four listed authors and consistently frames the performance figures as author-reported results, not established facts.

Authors:Hilaf Hasson, Aditya Chakravarty, Jayant Thomas, Krishna Gogineni

summary — pass: The title, summary and question faithfully convey the abstract's capability-cost gap argument, the query-time workspace and both system variants. Terms such as ingestion, vector index and judge score are left to tooltips, but the core idea remains understandable to a general reader: the system is cheap to set up and can count or combine across documents. The question is a fair unresolved issue rather than a leading claim.

ScalableRAG closes this gap with a persistent typed-set workspace: an agent builds named document and value sets at query time

date — pass: originalPublished 2026-07-27 matches the v1 submission date. The later v2 revision and the feed date are correctly not used.

[Submitted on 27 Jul 2026 (v1), last revised 26 Oct 2026 (this version, v2)]

support — pass: Every numeric claim matches the abstract: 71.74 on Hotels-50, 38.4 without the workspace, above 80 at 5M documents, 13 to 16 seconds per question, and under 20 minutes of ingestion. The scaling statement reflects the authors' own million-document viability claim. The limitation correctly bounds benchmark results without inventing stakeholders or impacts.

ScalableRAG combines minimal ingestion, conventional-QA parity, exact corpus operations, and million-document viability; no evaluated alternative provides all four.

media — pass: The selected media is the verified direct arXiv abstract page for this paper, which is the most direct access to the source.