GraphRAG: Structured Retrieval Augmented Generation (Microsoft)

Core Innovation

GraphRAG departs from traditional semantic-search RAG by constructing knowledge graphs from raw text, establishing community hierarchies, and generating community summaries to enhance language model reasoning.

Problem Statement

Baseline RAG approaches struggle in two critical scenarios: 1. Connecting disparate information pieces requiring synthesis across shared attributes 2. Providing holistic understanding of summarized concepts across large datasets

Competitive Advantage

Demonstrates "substantial improvement in answering" complex questions compared to vector-similarity-based retrieval methods, particularly for private datasets.

Process

Indexing Phase: - Converts input corpus into analyzable TextUnits - Extracts entities, relationships, and claims using LLMs - Applies hierarchical clustering via Leiden technique - Generates bottom-up community summaries

Query Modes: - Global Search: answers corpus-wide questions using community summaries - Local Search: explores specific entities and neighbors - DRIFT Search: combines local exploration with community context - Basic Search: fallback to traditional vector retrieval

Notable Points

"Using GraphRAG with your data out of the box may not yield the best possible results" -- recommends prompt tuning optimization. The project requires configuration updates between releases.