RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval

By Parth Sarthi, Salman Abdullah, Aditi Tuli, Shubh Khanna, Anna Goldie, Christopher D. Manning. International Conference on Learning Representations, 2024

578 citations recorded by the source corpus.

Retrieval-augmented language models can better adapt to changes in world state and incorporate long-tail knowledge. However, most existing methods retrieve only short contiguous chunks from a retrieval corpus, limiting holistic understanding of the overall document context. We introduce the novel approach of recursively embedding, clustering, and summarizing chunks of text, constructing a tree with differing levels of summarization from the bottom up. At inference time, our RAPTOR model retrieves from this tree, integrating information across lengthy documents at different levels of abstraction. Controlled experiments show that retrieval with recursive summaries offers significant improvements over traditional retrieval-augmented LMs on several tasks. On question-answering tasks that involve complex, multi-step reasoning, we show state-of-the-art results; for example, by coupling RAPTOR retrieval with the use of GPT-4, we can improve the best performance on the QuALITY benchmark by 20% in absolute accuracy.

Subfields: Artificial Intelligence, Signal Processing

View Paper: https://www.semanticscholar.org/paper/1d40f07569482d24e3335ffb31394d4bf1567e6f

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

1. Apurva N. Mody
2. Ha Trinh
3. Steven Truitt
4. Newman Cheng
5. A. Chao
6. Darren Edge
7. Joshua Bradley
8. Jonathan Larson

Citations in the corpus

• Retrieval-Augmented Generation for AI-Generated Content: A Survey (2024, unknown): “This paradigm makes the model learn to leverage API tools.”
• From Local to Global: A Graph RAG Approach to Query-Focused Summarization (2024, background): “…also bears resemblance to further approaches, such as generating a hierarchical index of text chunks by clustering the vectors of text embeddings (RAPTOR, Sarthi et al., 2024) or generating a “tree of clarifications” to answer multiple interpretations of ambiguous questions (Kim et al., 2023).”
• LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory (2024, unknown): “By comparsion, HippoRAG builds an entity-centric index, and RAPTOR/Memwalker build a hierarchical index using recursive summarization.”
• A Matryoshka Hierarchical RAG for Efficient Multi-Hop Question Answering (2026, unknown): “RAP-TOR [30] builds the tree recursively, representing each internal node with an LLM-generated summary of its child nodes.”
• MM-BizRAG: Rethinking Multimodal Retrieval-Augmented Generation for General Purpose Enterprise Q&A (2026, unknown): “Yet these pipelines often lack standardization (Gao et al., 2025; Zhang et al., 2025) in document representation (Sarthi et al., 2024; Jin et al., 2025; Yu et al., 2024a) and ingestion design (Xiong et al., 2024; Asai et al., 2023).”
• ASTRA-QA: A Benchmark for Abstract Question Answering over Documents (2026, unknown): “Recent RAG methods [2–4, 7, 24, 25] have moved further toward long-form and document-level synthesis through hierarchical retrieval, graph-structured indexing, and long-context reading.”
• MG2-RAG: Multi-Granularity Graph for Multimodal Retrieval-Augmented Generation (2026, unknown): “To overcome the structural limitations of flat retrieval, recent work has incorporated graph-based representations into RAG [16,28,53,62].”
• MDKeyChunker: What Does One LLM Call per Chunk Buy for Markdown Retrieval? (2026, unknown): “RAPTOR [Sarthi et al., 2024] splits text into short fixed-size leaf chunks (about 100 tokens), clusters their embeddings, summarizes each cluster with an LLM, and repeats on the summaries; leaves are kept and summary nodes are added, so the number of LLM calls is below one per leaf.”
• Mindscape-Aware Retrieval Augmented Generation for Improved Long Context Understanding (2025, unknown): “We compare against Qwen baselines and state-of-the-art RAG methods, including HippoRAG v2 (Guti’errez et al., 2025), MemoRAG (Qian et al., 2025) and RAPTOR (Sarthi et al., 2024).”
• Context is Gold to find the Gold Passage: Evaluating and Training Contextual Document Embeddings (2025, unknown): “For example, Edge et al. (2024) and Sarthi et al. (2024) cluster related chunks into semantic graphs or tree hierarchies.”
• ELITE: Embedding-Less retrieval with Iterative Text Exploration (2025, unknown): “More recent advances, such as MiniRAG [5] and Raptor [18], simplify retrieval pipelines by replacing static, large-scale graphs with lighter-weight structures or recursive chunk hierarchies.”
• RAG vs. GraphRAG: A Systematic Evaluation and Key Insights (2025, unknown): “…abstractions; (3) Text-centric graph-guided RAG [16], which retrieves original text chunks with the assistance of a constructed knowledge graph; and (4) Hierarchical summary-based GraphRAG [32], which builds hierarchical summaries to enable multi-granular retrieval without relying on explicit KGs.”
• ArchRAG: Attributed Community-based Hierarchical Retrieval-Augmented Generation (2025, unknown): “As shown in the literature, a few recent graph-based RAG methods have emerged for both abstract question [12, 20] and specific question [20, 21, 34, 54, 69, 70], where the former targets high-level themes (e.g., “What are the potential impacts of LLMs on education?”)”