RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval

PAPER-DERIVED FACTS: RAPTOR constructs a tree of document representations by recursively embedding, clustering, and summarizing chunks. Retrieval can then select information at different levels of abstraction rather than only from short contiguous chunks.

PAPER-DERIVED FACTS: Flat chunk retrieval is effective for local facts but can lose relationships and themes spanning a long document. RAPTOR introduces hierarchical summaries so retrieval can operate over both detailed and abstract representations.

PAPER-DERIVED FACTS: The system must answer questions whose evidence may be distributed across distant sections of a document. A flat retriever may retrieve isolated chunks without the broader context needed for multi-step reasoning.

PAPER-DERIVED FACTS: RAPTOR builds on dense embeddings, clustering, abstractive summarization, and retrieval-augmented generation. Its main distinction is that summaries are recursively organized into a tree.

PAPER-DERIVED FACTS: Leaf chunks are embedded and clustered; an LLM summarizes each cluster; the summaries are embedded and clustered again; this repeats until higher-level representations are formed. At inference, retrieval can traverse or collapse the tree.

PAPER-DERIVED FACTS: The index is a multi-level tree containing original chunks at the leaves and progressively more abstract summaries above them. A query embedding is used to retrieve relevant nodes, which are then supplied to the language model.

PAPER-DERIVED FACTS: The evaluation covers long-document question answering and benchmarks requiring multi-step reasoning, including QuALITY and other QA datasets reported in the paper.

PAPER-DERIVED FACTS: RAPTOR does not require end-to-end retraining of the generator. Index construction uses embedding models and an LLM summarizer; the resulting tree can be paired with a downstream language model such as GPT-4.

PAPER-DERIVED FACTS: Experiments compare tree-based retrieval with standard chunk retrieval and evaluate both retrieval configurations and final QA. The authors test recursive summaries on questions requiring information from different parts of documents.

PAPER-DERIVED FACTS: Baselines include conventional flat retrieval systems that select contiguous chunks. The comparison tests whether hierarchical context provides a measurable advantage beyond simply retrieving more chunks.

PAPER-DERIVED FACTS: RAPTOR improves performance on several long-document QA tasks and reports a 20 percentage-point absolute improvement on QuALITY when combined with GPT-4 over the previous best result cited by the authors.

PAPER-DERIVED FACTS: The work compares tree traversal retrieval with collapsed-tree retrieval and studies the effect of hierarchical summaries. These experiments examine how abstraction level and retrieval strategy influence answer quality.

PAPER-DERIVED FACTS: Summaries can lose fine-grained details, while leaf chunks can lack global context. RAPTOR therefore faces a granularity trade-off: abstract nodes improve holistic understanding but may omit exact evidence.

PAPER-DERIVED FACTS: RAPTOR turns a flat document collection into a hierarchy of information at different abstraction levels. This enables RAG systems to retrieve both local facts and global summaries from long documents.

PAPER-DERIVED FACTS: Primary source: Sarthi, P. et al. (2024), RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval. ICLR 2024. arXiv:2401.18059.