TTT #44: Improving Summarization Tasks with GraphRAG and RaptorRAG

Stephen Collins • May 18, 2024

In this issue, I’ll look at two new methods for improving how we find relevant information (the “context”) for tasks where large language models (LLMs) summarize text: GraphRAG and RaptorRAG. These methods help us create more detailed and varied answers to difficult questions using large amounts of text.

GraphRAG: Leveraging Graphs for Global Summarization

GraphRAG, or Graph-based Retrieval-Augmented Generation, is designed to handle global questions that require summarizing information across entire datasets. Traditional RAG systems struggle with these questions as they focus on retrieving local chunks of text.

In the paper “From Local to Global: A Graph RAG Approach to Query-Focused Summarization,” Darren Edge et al. introduce a two-stage pipeline. An LLM generates an entity knowledge graph from source documents, capturing nodes (entities), edges (relationships), and covariates (claims). Community detection algorithms partition the graph into closely-related groups, which are summarized independently. At query time, these community summaries generate partial responses, which are then synthesized into a final global answer.

In my blog post, “Implementing GraphRAG for Query-Focused Summarization,” I provide a step-by-step guide to building an entity knowledge graph and generating community summaries using Python. This tutorial is aimed at intermediate developers looking to leverage advanced LLM capabilities in their applications.

RaptorRAG: Recursive Abstraction for Enhanced Retrieval

RaptorRAG (Recursive Abstractive Processing for Tree-Organized Retrieval) takes a different approach to improve retrieval-augmented generation. Traditional methods retrieve short, contiguous text chunks, which often fail to capture the broader context needed for complex queries.

Detailed in the paper “RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval” by Parth Sarthi et al., RaptorRAG builds a hierarchical tree structure by recursively embedding, clustering, and summarizing text chunks. This tree structure captures both high-level and low-level details, allowing for more effective retrieval and integration of information across lengthy documents. Controlled experiments show significant improvements in performance on complex, multi-step reasoning tasks.

Comparing GraphRAG and RaptorRAG

While both GraphRAG and RaptorRAG enhance query-focused summarization, they differ fundamentally in their approaches:

  • Structure: GraphRAG uses an entity knowledge graph to organize information, emphasizing relationships and community detection. RaptorRAG, on the other hand, builds a hierarchical tree by clustering and summarizing text chunks, focusing on multi-level abstraction.
  • Summarization: GraphRAG generates community summaries based on detected relationships, while RaptorRAG produces recursive summaries at different levels of abstraction.
  • Query Handling: GraphRAG’s community-based approach is tailored for global summarization, ensuring comprehensive and diverse answers. RaptorRAG excels in integrating information from various document parts, enhancing retrieval for complex, multi-step queries.

Integrating GraphRAG and RaptorRAG

Combining GraphRAG and RaptorRAG could create powerful systems for query-focused summarization. Using RaptorRAG’s hierarchical retrieval with GraphRAG’s community detection and summarization offers a robust solution for handling complex queries over large datasets.

Conclusion

GraphRAG and RaptorRAG represent significant advancements in retrieval-augmented generation. By leveraging these innovative methods, we can generate detailed and diverse answers to complex queries. I encourage you to explore these approaches and consider how they can be integrated into your projects.