---
source: https://iinventedlight.com/library/notes/ao-context-retrieval-trees.md
title: "Retrieval trees: summarizing chunks into levels (RAPTOR)"
license: CC-BY-4.0
license_url: https://creativecommons.org/licenses/by/4.0/
source_license: "CC BY 4.0; restated in our own words"
attribution: "Summary by iinventedlight (AI-assisted). Sources: arXiv, [2401.18059] RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval, CC BY 4.0; arXiv, [2402.06196] Large Language Models: A Survey, CC BY 4.0."
author: "iinventedlight (synthesis of cited sources); text drafted with AI assistance"
date: 2026-11-04
status: published
checked_on: 2026-11-04
as_of: 2026-11-04
confidence: medium
untested: "No experiment was run; only the fetched arXiv abstract pages were read, so results are as reported by the authors."
category: agent-operations
topic_area: "agent-ops: context and retrieval"
tags: ['agent-operations', 'retrieval', 'rag', 'context']
sources:
  - id: S1
    title: "[2401.18059] RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval"
    publisher: arXiv
    url: https://arxiv.org/abs/2401.18059
    accessed: 2026-11-04
    license: "CC BY 4.0"
  - id: S2
    title: "[2402.06196] Large Language Models: A Survey"
    publisher: arXiv
    url: https://arxiv.org/abs/2402.06196
    accessed: 2026-11-04
    license: "CC BY 4.0"
---
# Retrieval trees: summarizing chunks into levels (RAPTOR)

**Answer:** Instead of retrieving only short, flat chunks, a retrieval tree recursively embeds, clusters and summarizes chunks so a model can pull context at several levels of detail. The RAPTOR paper reports this helps on long, multi-step question answering.

## What the method does

Plain retrieval-augmented generation (RAG) splits documents into small contiguous chunks and fetches the closest few. The RAPTOR authors argue this limits understanding of the whole document. Their approach builds a tree bottom-up: leaf chunks are embedded and clustered, each cluster is summarized, and the summaries are clustered and summarized again. At query time the model can retrieve from any level, mixing fine detail with high-level summaries.

## Reported results

The authors report significant gains over traditional retrieval-augmented models on several tasks, and state-of-the-art results on complex multi-step question answering. Paired with GPT-4, they report a 20 point absolute accuracy improvement on the QuALITY benchmark.

## Where it fits

A broad LLM survey (S2) places retrieval augmentation among the main techniques used to build and augment LLMs, alongside training data and evaluation work. Retrieval trees are one way to scale that idea to long documents without stuffing the whole text into the context window.

## Practical reading

- Tree building has an upfront cost: every summary level needs model calls.
- It suits long, stable corpora (manuals, books, reports) more than fast-changing data.
- Results above are the authors' own benchmarks, not independent replications.

## Claims
- id: C1
  claim: RAPTOR recursively embeds, clusters and summarizes text chunks into a tree.
  source: S1
  evidence: "recursively embedding, clustering, and summarizing chunks of text, constructing a tree"
- id: C2
  claim: Most existing methods retrieve only short contiguous chunks.
  source: S1
  evidence: "most existing methods retrieve only short contiguous chunks from a retrieval corpus"
- id: C3
  claim: With GPT-4, RAPTOR improved QuALITY accuracy by 20 points absolute.
  source: S1
  evidence: "we can improve the best performance on the QuALITY benchmark by 20% in absolute accuracy"
- id: C4
  claim: The survey reviews techniques developed to build and augment LLMs.
  source: S2
  evidence: "We also give an overview of techniques developed to build, and augment LLMs."

## Sources
- [S1] [2401.18059] RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval. arXiv. https://arxiv.org/abs/2401.18059 . License: CC BY 4.0. Accessed 2026-11-04.
- [S2] [2402.06196] Large Language Models: A Survey. arXiv. https://arxiv.org/abs/2402.06196 . License: CC BY 4.0. Accessed 2026-11-04.

License: this note is CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Summary by iinventedlight (AI-assisted).