RAPTOR is a tree-based retrieval system designed to enhance search capabilities for language models.
It clusters text chunks and recursively generates summaries to build a tree structure, allowing for efficient retrieval of information at different levels of granularity.
RAPTOR outperforms traditional retrieval methods in question-answering tasks, particularly in complex reasoning and multi-hop questions.
When combined with GPT-4, RAPTOR achieves state-of-the-art results, improving accuracy on the QuALITY benchmark by 20%.