MachinoAI explainer / Retrieval-Augmented Generation
From Local to Global: A Graph RAG Approach to Query-Focused Summarization
Introduces GraphRAG, which uses entity graphs and community summaries to answer global questions that flat chunk retrieval struggles with.
01
Abstract
PAPER-DERIVED FACTS: GraphRAG combines retrieval-augmented generation with a graph-based index for questions that require understanding an entire corpus. It extracts entities and relationships, detects communities, generates community summaries, and synthesizes partial answers into a global response.
02
Introduction
PAPER-DERIVED FACTS: Traditional RAG is effective for local questions that map to a few relevant chunks, but global questions such as identifying themes across a corpus require aggregation. GraphRAG introduces an index designed for corpus-level sensemaking.
03
Problem
PAPER-DERIVED FACTS: The challenge is to answer broad questions over large private collections without feeding the entire corpus to an LLM. The system must preserve relationships and thematic structure while reducing the amount of context processed at query time.
04
Background
PAPER-DERIVED FACTS: GraphRAG builds on knowledge graph construction, community detection, summarization, and retrieval-augmented generation. Its design is related to query-focused summarization but adds a graph index that scales the source corpus.
05
Methodology
PAPER-DERIVED FACTS: Documents are split into text units, an LLM extracts entities and relationships, and a graph is constructed. Community detection groups related entities, and an LLM generates summaries for each community; higher-level summaries can be built recursively.
Figure notes
Visual evidence
06
Architecture
PAPER-DERIVED FACTS: The indexing path transforms source documents into text units, entities and relationships, a knowledge graph, communities, and community summaries. Global querying generates partial answers from selected community reports and combines them into a final answer.
07
Dataset
PAPER-DERIVED FACTS: The paper evaluates global sensemaking on private-text corpora in the roughly one-million-token range and uses generated questions designed to test comprehensiveness and diversity of corpus-level answers.
08
Training
PAPER-DERIVED FACTS: GraphRAG is primarily an indexing and inference architecture rather than a newly trained foundation model. LLMs are used during indexing for entity extraction and summarization and at query time for answer synthesis.
09
Experiments
PAPER-DERIVED FACTS: The evaluation compares GraphRAG with a conventional RAG baseline on global questions. Metrics focus on the comprehensiveness and diversity of generated answers, where local retrieval can omit broad themes.
10
Baselines
PAPER-DERIVED FACTS: The main baseline is a naive RAG pipeline that retrieves local text chunks using conventional vector search. The comparison tests whether graph-derived community summaries improve corpus-level answers.
11
Results
PAPER-DERIVED FACTS: For the studied global sensemaking questions, GraphRAG shows substantial improvements over naive RAG in both comprehensiveness and diversity of generated answers.
12
Ablation
PAPER-DERIVED FACTS: The system can support different retrieval strategies, including global community-summary retrieval and local graph-based retrieval. The architecture highlights a trade-off between broad thematic coverage and precise entity-level evidence.
13
Limitations
Error Analysis
PAPER-DERIVED FACTS: Graph construction and summarization can introduce extraction or abstraction errors. Community summaries may omit fine details, while global synthesis can become less precise when too many community-level answers are combined.
14
Conclusion
PAPER-DERIVED FACTS: GraphRAG extends RAG with an offline graph-and-summary index designed for global questions. Its central insight is to retrieve structured summaries of communities rather than isolated text chunks when the query requires corpus-level understanding.
15
References
PAPER-DERIVED FACTS: Primary source: Edge, D. et al. (2024), From Local to Global: A Graph RAG Approach to Query-Focused Summarization. arXiv:2404.16130.
Continue reading
Related research
Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
Defines the canonical RAG architecture that combines a dense retriever, an external document index, and a seq2seq generator trained end-to-end.
Retrieval-Augmented GenerationDense Passage Retrieval for Open-Domain Question Answering
Introduces a practical dual-encoder dense retriever that replaces sparse lexical matching with learned semantic passage representations.
Retrieval-Augmented GenerationREALM
Introduces end-to-end retrieval during language-model pre-training so factual knowledge can live in an external, replaceable corpus rather than only in model parameters.