RAPTOR is a novel tree-based retrieval system that enhances the contextual understanding of large language models (LLMs) by recursively clustering and summarizing text chunks. The system constructs a hierarchical tree structure from the bottom up, enabling efficient retrieval of information at different levels of abstraction. This approach addresses the limitation of traditional retrieval-augmented models, which often retrieve only short, contiguous text chunks, leading to incomplete understanding of long documents. RAPTOR's tree structure allows for more comprehensive retrieval by integrating information across different levels of abstraction, improving performance on complex, multi-step reasoning tasks. Controlled experiments show that RAPTOR outperforms existing retrieval-augmented models on several tasks, including question-answering, with significant improvements in accuracy. When combined with GPT-4, RAPTOR achieves a 20% absolute accuracy improvement on the QuALITY benchmark. The system's tree-based retrieval mechanism, which includes both tree traversal and collapsed tree strategies, enables efficient and effective retrieval of information from large text corpora. RAPTOR's approach is scalable and efficient, with linear scaling in terms of build time and token expenditure, making it suitable for processing large and complex documents. The system's effectiveness is demonstrated through experiments on three question-answering datasets: NarrativeQA, QASPER, and QuALITY, where RAPTOR consistently outperforms traditional retrieval methods. The results highlight the importance of hierarchical information retrieval in capturing both broad themes and detailed information, leading to improved performance on complex tasks.RAPTOR is a novel tree-based retrieval system that enhances the contextual understanding of large language models (LLMs) by recursively clustering and summarizing text chunks. The system constructs a hierarchical tree structure from the bottom up, enabling efficient retrieval of information at different levels of abstraction. This approach addresses the limitation of traditional retrieval-augmented models, which often retrieve only short, contiguous text chunks, leading to incomplete understanding of long documents. RAPTOR's tree structure allows for more comprehensive retrieval by integrating information across different levels of abstraction, improving performance on complex, multi-step reasoning tasks. Controlled experiments show that RAPTOR outperforms existing retrieval-augmented models on several tasks, including question-answering, with significant improvements in accuracy. When combined with GPT-4, RAPTOR achieves a 20% absolute accuracy improvement on the QuALITY benchmark. The system's tree-based retrieval mechanism, which includes both tree traversal and collapsed tree strategies, enables efficient and effective retrieval of information from large text corpora. RAPTOR's approach is scalable and efficient, with linear scaling in terms of build time and token expenditure, making it suitable for processing large and complex documents. The system's effectiveness is demonstrated through experiments on three question-answering datasets: NarrativeQA, QASPER, and QuALITY, where RAPTOR consistently outperforms traditional retrieval methods. The results highlight the importance of hierarchical information retrieval in capturing both broad themes and detailed information, leading to improved performance on complex tasks.