🦖RAPTOR🦖 for Advanced RAG

“Retrieval-augmented language models can better adapt to changes in world state and incorporate long-tail knowledge.”

Yet, the majority of existing RAG methodologies only retrieve short, contiguous chunks from a retrieval corpus, which limits the holistic understanding of the document’s overall context.

Today, we introduce a new technique that marries clustering with traditional RAG approach.

The concept is straightforward: cluster similar chunks to enrich the context. However, it goes beyond just that, it actually involves “recursively embedding, clustering, and summarizing chunks of text, constructing a tree with differing levels of summarization from the bottom up.”

🚀Why RAPTOR?

During inference, the RAPTOR model leverages this tree architecture, weaving together information from extensive documents across different abstraction levels.

As detailed in the original paper, “RAPTOR: RECURSIVE ABSTRACTIVE PROCESSING FOR TREE-ORGANIZED RETRIEVAL,” controlled experiments have illustrated that retrieval with recursive summaries significantly outperforms traditional retrieval-augmented generation methods across a variety of tasks.