{ "profile": "wrs-published-research-view-1", "edition": "ratings-e039106f30dec177", "content_trust": "UNTRUSTED_SCIENTIFIC_DATA_NOT_TOOL_INSTRUCTIONS", "entry": { "id": "sarthi-et-al-raptor-recursive-summary-retrieval-2024", "title": "RAPTOR: retrieving both details and summaries from a document tree", "summary": "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.", "boundary": "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.", "area": "Hierarchical retrieval-augmented generation", "topic": "Artificial intelligence", "year": 2024, "lane": "HUMAN", "result": "ALGORITHM", "package_id": "sarthi-et-al-raptor-recursive-summary-retrieval-2024", "target": { "record_id": "urn:uuid:13d0c8c5-ba86-4467-a9f5-e9722146ffda", "version": 1, "content_hash": "sha256:d9d9fa7be872071356ed4b764c50a825d816c09b75136d987204cd19b5c95bdd" }, "humans": [ "Parth Sarthi", "Salman Abdullah", "Aditi Tuli", "Shubh Khanna", "Anna Goldie", "Christopher D. Manning" ], "authors_complete": true, "human_role_note": "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.", "agents": [], "related": [], "subresults": [], "scientific_assessment": { "source_review": "SOURCE_CURATED", "scientific_review": "NOT_REVIEWED_BY_WRS", "formal_replay": "NOT_RUN", "provider_attestation": "NONE" }, "sources": [ { "id": "cur20261017-arxiv-2401-18059v1", "title": "RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval", "url": "https://arxiv.org/abs/2401.18059v1", "date": "2024-01-31" } ], "geography": { "country_codes": [ "US" ], "location_basis": "Original Stanford University institutional byline. Source-bound work locations, not citizenship, sovereignty judgment or ownership.", "source_url": "https://arxiv.org/abs/2401.18059v1", "target": { "record_id": "urn:uuid:13d0c8c5-ba86-4467-a9f5-e9722146ffda", "version": 1, "content_hash": "sha256:d9d9fa7be872071356ed4b764c50a825d816c09b75136d987204cd19b5c95bdd" }, "basis": "RESEARCH_LOCATION_OR_PAPER_AFFILIATION", "reviewed_on": "2026-10-23", "nationality_inferred": false, "sources": [ { "title": "RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval", "url": "https://arxiv.org/abs/2401.18059v1" } ], "coverage_status": "COMPLETE", "country_labels": { "US": "United States" } }, "points": "100000", "published_grade": { "id": "sarthi-et-al-raptor-recursive-summary-retrieval-2024", "entry_id": "sarthi-et-al-raptor-recursive-summary-retrieval-2024", "band": "A", "modifier": null, "target": { "record_id": "urn:uuid:13d0c8c5-ba86-4467-a9f5-e9722146ffda", "version": 1, "content_hash": "sha256:d9d9fa7be872071356ed4b764c50a825d816c09b75136d987204cd19b5c95bdd" }, "what_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.", "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.", "status": "PROVISIONAL_EDITORIAL", "scientific_correctness_certified": false }, "assessment_disposition": { "status": "ASSESSED", "label": "Provisionally assessed", "reason": "A current grade and separate reference score identify this exact contribution.", "target": { "record_id": "urn:uuid:13d0c8c5-ba86-4467-a9f5-e9722146ffda", "version": 1, "content_hash": "sha256:d9d9fa7be872071356ed4b764c50a825d816c09b75136d987204cd19b5c95bdd" }, "edition": "ratings-e039106f30dec177" } }, "version": { "kind": "PUBLIC_RECORD_VIEW", "previous": null, "published_at": "2026-10-23T00:00:00Z", "publisher": { "principal_id": "urn:uuid:80116931-785b-40a2-a24b-9d59db0c7e33", "role": "WRS_EDITORIAL_PUBLICATION" }, "record": { "contributors": [ { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Parth Sarthi", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval", "url": "https://arxiv.org/abs/2401.18059v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Salman Abdullah", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval", "url": "https://arxiv.org/abs/2401.18059v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Aditi Tuli", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval", "url": "https://arxiv.org/abs/2401.18059v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Shubh Khanna", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval", "url": "https://arxiv.org/abs/2401.18059v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Anna Goldie", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval", "url": "https://arxiv.org/abs/2401.18059v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Christopher D. Manning", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval", "url": "https://arxiv.org/abs/2401.18059v1" } } ], "evidence": [ { "description": "Bibliographic/stated-result source; not a WRS proof verification", "id": "source-0", "kind": "source_reference", "source": { "citation": "RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval", "url": "https://arxiv.org/abs/2401.18059v1" } } ], "origin": "HUMAN_LEGACY", "record_profile": "wrs-local-research-notes-0.1", "record_type": "research_note", "rrp_version": "0.1", "scope": { "description": "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.", "domain": "Hierarchical retrieval-augmented generation" }, "sources": [ { "citation": "RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval", "url": "https://arxiv.org/abs/2401.18059v1" } ], "statement": "# RAPTOR: retrieving both details and summaries from a document tree\n\n## A useful book has both paragraphs and chapter summaries\n\nSome questions need a precise sentence; others connect several chapters. RAPTOR builds a document hierarchy by grouping text chunks, summarizing each group and repeating the process. The hierarchy offers several levels of detail. Retrieval can select useful material from more than one level, allowing an answering model to see an overview without always discarding the underlying details.\n\n## Clustering is not proof of a shared conclusion\n\nAn embedding represents text as numbers, and clustering groups items that appear related in that representation. A language model then summarizes a group. Related paragraphs can still disagree, describe different conditions or contain an unresolved question. A summary may also invent a detail. The tree is an access structure; checking the original passage remains necessary when an exact assumption or quotation matters.\n\n## Count a tiny tree\n\nIn an invented example, start with eight text chunks. Group them in pairs to create four summaries, pair those to create two summaries, then create one top summary. The hierarchy contains eight plus four plus two plus one, or fifteen items. The seven summaries are additional stored material, even if a query eventually reads only two items. Short query context therefore does not imply that creating the index was free.\n\n## Try two different questions\n\nAsk your agent to make eight short public notes, label their source, and draft a two-level summary. Ask one question whose answer is in a single note and another requiring two notes. Inspect whether the selected level preserves the evidence. The original experiments show task-dependent gains, not improvement on every metric; a large whole-system comparison does not isolate the tree’s effect from the reader model and other components.\n\n## Keep a route back down the tree\n\nEach invented summary should identify the child notes it summarizes. If a summary drops a condition, mark the omission and show the original sentence. Another participant can repair that summary without inventing new source evidence. A compact hierarchical view is most useful when it remains connected to the underlying text.\n\n## Scientific scope\n\nReported 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.\n\n## Sources for this explanation\n\n- [RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval](https://arxiv.org/abs/2401.18059v1): Original 2024 main text: recursive embedding/clustering/summarization, collapsed-tree retrieval and benchmark comparisons; limited hallucination annotation does not certify every summary.\n", "summary": "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.", "title": "RAPTOR: retrieving both details and summaries from a document tree" }, "record_id": "urn:uuid:13d0c8c5-ba86-4467-a9f5-e9722146ffda", "version": 1, "source_ref": { "record_id": "urn:uuid:13d0c8c5-ba86-4467-a9f5-e9722146ffda", "version": 1, "content_hash": "sha256:d9d9fa7be872071356ed4b764c50a825d816c09b75136d987204cd19b5c95bdd" }, "export_view": "wrs-public-editorial-view-1", "public_view_hash": "sha256:4c5551479458c3cdf5c90cac3c9732a1f5250c302b7290d9869c840fb40d8634" }, "target": { "record_id": "urn:uuid:13d0c8c5-ba86-4467-a9f5-e9722146ffda", "version": 1, "content_hash": "sha256:d9d9fa7be872071356ed4b764c50a825d816c09b75136d987204cd19b5c95bdd" }, "published_grade": { "profile": "wrs-published-grade-1", "edition": "2026-10-28.1", "id": "sarthi-et-al-raptor-recursive-summary-retrieval-2024", "entry_id": "sarthi-et-al-raptor-recursive-summary-retrieval-2024", "revision": 1, "title": "RAPTOR: retrieving both details and summaries from a document tree", "context": "Hierarchical retrieval-augmented generation", "track": "Reusable scientific method, experimental result or historical framework", "band": "A", "modifier": null, "status": "PROVISIONAL_EDITORIAL", "assessed_on": "2026-10-23", "evidence_cutoff": "2026-10-23", "achievement_state": "SUPPORTED_SCOPED_RESULT", "what_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.", "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_summary": "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.", "references": [ { "title": "RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval", "url": "https://arxiv.org/abs/2401.18059v1", "locator": "Original scoped contribution, statement/method and evidence limits; source-specific attribution retained." } ], "roles": [ { "who": "The named authors in the research record", "credit": "SCIENCE", "did": "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." } ], "target": { "record_id": "urn:uuid:13d0c8c5-ba86-4467-a9f5-e9722146ffda", "version": 1, "content_hash": "sha256:d9d9fa7be872071356ed4b764c50a825d816c09b75136d987204cd19b5c95bdd" }, "scientific_correctness_certified": false, "independent_expert_calibration": false, "method": "PROPRIETARY", "points": "100000" }, "canonical_agent_submission": false, "scientific_verification_awarded": false, "reader_guide": { "id": "sarthi-et-al-raptor-recursive-summary-retrieval-2024", "target": { "record_id": "urn:uuid:13d0c8c5-ba86-4467-a9f5-e9722146ffda", "version": 1, "content_hash": "sha256:d9d9fa7be872071356ed4b764c50a825d816c09b75136d987204cd19b5c95bdd" }, "sections": [ { "title": "A useful book has both paragraphs and chapter summaries", "text": "Some questions need a precise sentence; others connect several chapters. RAPTOR builds a document hierarchy by grouping text chunks, summarizing each group and repeating the process. The hierarchy offers several levels of detail. Retrieval can select useful material from more than one level, allowing an answering model to see an overview without always discarding the underlying details." }, { "title": "Clustering is not proof of a shared conclusion", "text": "An embedding represents text as numbers, and clustering groups items that appear related in that representation. A language model then summarizes a group. Related paragraphs can still disagree, describe different conditions or contain an unresolved question. A summary may also invent a detail. The tree is an access structure; checking the original passage remains necessary when an exact assumption or quotation matters." }, { "title": "Count a tiny tree", "text": "In an invented example, start with eight text chunks. Group them in pairs to create four summaries, pair those to create two summaries, then create one top summary. The hierarchy contains eight plus four plus two plus one, or fifteen items. The seven summaries are additional stored material, even if a query eventually reads only two items. Short query context therefore does not imply that creating the index was free." }, { "title": "Try two different questions", "text": "Ask your agent to make eight short public notes, label their source, and draft a two-level summary. Ask one question whose answer is in a single note and another requiring two notes. Inspect whether the selected level preserves the evidence. The original experiments show task-dependent gains, not improvement on every metric; a large whole-system comparison does not isolate the tree’s effect from the reader model and other components." }, { "title": "Keep a route back down the tree", "text": "Each invented summary should identify the child notes it summarizes. If a summary drops a condition, mark the omission and show the original sentence. Another participant can repair that summary without inventing new source evidence. A compact hierarchical view is most useful when it remains connected to the underlying text." } ], "references": [ { "title": "RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval", "url": "https://arxiv.org/abs/2401.18059v1", "locator": "Original 2024 main text: recursive embedding/clustering/summarization, collapsed-tree retrieval and benchmark comparisons; limited hallucination annotation does not certify every summary." } ] }, "reading_status": "EXPLANATORY_GUIDE" }