From Local to Global: Mastering Query-Focused Summarization with GraphRAG

GraphRAG represents an innovative technique enhancing the capabilities of retrieval-augmented generation (RAG) systems. Traditional RAG systems efficiently retrieve specific pieces of information from large datasets to answer localized queries. However, they struggle with global queries that require summarization across entire collections. GraphRAG addresses this challenge by integrating graph-based indexing with summarization processes, enabling large language models (LLMs) to produce comprehensive, detailed, and diverse answers to broad queries.

The GraphRAG Approach

GraphRAG operates in two main stages: graph-based text indexing and community summarization. Initially, an LLM processes the source documents to create an entity knowledge graph, identifying nodes (entities) and edges (relationships) among them. This graph is then partitioned into modular communities using algorithms like Leiden, which cluster closely related nodes together. Each community undergoes summarization to generate partial responses that are subsequently merged to form a final, comprehensive answer.

Advantages of GraphRAG

The primary advantage of GraphRAG lies in its ability to handle extensive datasets and produce detailed, high-quality summaries for global queries. By leveraging the modularity of graphs, GraphRAG ensures that the summarization process covers all relevant aspects of the dataset, maintaining both the comprehensiveness and diversity of the generated answers. This approach also allows for efficient processing, as it partitions the dataset into manageable chunks that can be processed in parallel, optimizing the use of LLMs.

Using GraphRAG

1. Document Processing: Split the source documents into text chunks suitable for LLM processing.
2. Entity and Relationship Extraction: Use LLM prompts to identify and extract entities and relationships from the text chunks.
3. Graph Construction: Build a graph index from the extracted entities and relationships.
4. Community Detection: Apply community detection algorithms to partition the graph into clusters of related entities.
5. Summarization: Generate summaries for each community and combine these summaries to answer the query.

Step by step approach using GraphRAG Python Library: Get Started (microsoft.github.io)

GraphRAG in practice

I tried the GraphRAG approach by indexing the actual study: [2404.16130] From Local to Global: A Graph RAG Approach to Query-Focused Summarization (arxiv.org)

First I installed a GraphRAG library, initiated entity extraction and built a graph index:

Statistics of the index, communities and summaries:

The final step is to use the prompt global method to summarize the GraphRAG study:

Cost

The method is highly effective for building comprehensive summaries, however, it is important to take cost into a consideration. With 15 pages of text the building of the entities, index, communities and summaries + my prompt required clsoe to 300k tokens and resulted in cost of over 2$.

In conclusion, GraphRAG offers a significant advancement in query-focused summarization, combining the strengths of graph indexing and LLM summarization to handle global queries efficiently. This technique not only enhances the performance of LLMs but also ensures the production of detailed and diverse answers, making it invaluable for various applications, from scientific research to intelligence analysis.