AWhat changed
A question may need a precise paragraph or an overview spread across several sections. RAPTOR groups document chunks, summarizes the groups and repeats this process to build a hierarchy. Retrieval can consider text from different levels, rather than relying only on the original chunks. Sarthi and colleagues evaluate the design on several question-answering datasets. The contribution is this recursive clustering-and-summary retrieval method, building on existing embeddings, clustering and language models rather than counting each inherited component as a new discovery.
Scientific scope
A question may need a precise paragraph or an overview spread across several sections. RAPTOR groups document chunks, summarizes the groups and repeats this process to build a hierarchy. Retrieval can consider text from different levels, rather than relying only on the original chunks. Sarthi and colleagues evaluate the design on several question-answering datasets. The contribution is this recursive clustering-and-summary retrieval method, building on existing embeddings, clustering and language models rather than counting each inherited component as a new discovery. Reported gains depend on the task, reader model and comparison system. A large headline improvement compares whole systems, not an isolated causal effect of the tree. Some word-overlap metrics favor another baseline. Summaries can hallucinate, and a limited annotation study does not guarantee that errors never propagate. The retrieved-context budget excludes offline construction costs. The tree is a useful access structure, not a replacement for checking the original passage when an exact condition matters.
Evidence
The linked original works support the scoped statement and contribution attribution. WRS has not independently replayed experiments or reconstructed a full mathematical proof. Reported gains depend on the task, reader model and comparison system. A large headline improvement compares whole systems, not an isolated causal effect of the tree. Some word-overlap metrics favor another baseline. Summaries can hallucinate, and a limited annotation study does not guarantee that errors never propagate. The retrieved-context budget excludes offline construction costs. The tree is a useful access structure, not a replacement for checking the original passage when an exact condition matters.
Contributors
- The named authors in the research record · SCIENCE: Source-listed human researchers and engineers for the exact method or resource. The evaluated learning system is not an independently attested scientific author. Earlier methods, funding and data-collection roles remain separate. Individual effort shares and unresolved person identities are not inferred.
Provisional WRS assessment, edition 2026-10-23.9. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.
Full scientific record · Browse contribution grades · Request a correction