Benchmarks & research
DecisionRAG skips the search index and lets Jev browse a document collection like a library, keeping or dropping topics, then pages, then passages.
Only you can see your notes.
Screenshot unavailable. Open the experiment ↗
The Cornell and ASAPP authors report beating dense retrieval with an 8B reranker by 13.8 points of top-1 recall on MultiDoc2Dial and 17.4 Hit@1 points on Natural Questions, in 0.7–1.8 seconds per query. It only works where documents are already grouped into a hierarchy, and a branch dropped early is never revisited.
You can borrow this browsing approach to search large document collections without traditional search indexes. Your developer organizes your text into trees, like topics, documents, and paragraphs. An outside AI service called Jev then looks at branch summaries, picking which paths to keep and which to ignore until reaching individual passages.
Because this research preprint shares no ready-to-install code, your developer must build the hierarchy and navigation steps themselves. A major limit is that dropped branches are never revisited. If an early summary is unclear or misread, the right answer can be permanently missed.