RAPTOR

A hierarchical retrieval method — RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval (Sarthi, Abdullah, Tuli, Khanna, Goldie, Christopher D. Manning; Stanford, ICLR 2024; arXiv:2401.18059). It is grouped with the graph-RAG methods in graphrag-bench, but the important distinction is that RAPTOR is a tree, not an entity knowledge-graph: it gets multi-level summarization without OpenIE, entities, or community detection.

How it works

• Recursive clustering + summarization (bottom-up). Chunk the corpus, embed the chunks, cluster similar ones, and have an LLM write an abstractive summary of each cluster. Treat those summaries as a new layer and repeat — cluster the summaries, summarize again — until you reach a small set of top-level nodes. The result is a tree whose leaves are raw chunks and whose upper nodes are progressively more abstract summaries.
• Retrieval across levels. At query time RAPTOR pulls nodes from multiple tree levels at once (the “collapsed tree” mode searches all nodes regardless of depth), so an answer can mix a high-level summary with the specific leaf chunks under it — whole-document context and detail together, where plain chunk-RAG only ever returns short contiguous passages.

What it’s for

Long-document, multi-step, “what’s the overall point?” questions that span a whole text. Coupling RAPTOR retrieval with GPT-4 improved the best QuALITY result by +20% absolute accuracy, with strong showings on long-context QA (NarrativeQA, QASPER).

Where it sits in the cluster

RAPTOR (Jan 2024) predates Microsoft GraphRAG (Apr 2024) and reaches the same core insight — summarize clusters at multiple levels of abstraction so retrieval can answer global, not just local, questions — by a different route: a recursive summary tree of text chunks instead of an entity graph + Leiden communities. So it’s the tree cousin of GraphRAG’s community summaries, and a useful reminder that “graph”-RAG benchmarks like graphrag-bench really test structured/hierarchical retrieval broadly — of which a knowledge graph is one structure and a summary tree another. Distinct from HippoRAG2 (PPR over an entity+passage graph) and LightRAG (dual-level keyword retrieval). Authors/Stanford noted inline (thin).

Related

graphrag-bench · graphrag · from-local-to-global-graphrag · hipporag2 · lightrag · retrieval-augmented-generation · knowledge-graph