GraphRAG: Unlocking LLM discovery on narrative private data

Source: https://www.microsoft.com/en-us/research/blog/graphrag-unlocking-llm-discovery-on-narrative-private-data/ Publisher: Microsoft Research · Published: 13 February 2024 Authors: Jonathan Larson (Partner Data Architect), Steven Truitt (Principal Program Manager)

Summary

The originating write-up of GraphRAG: using LLM-generated knowledge graphs to answer questions over private data that baseline RAG cannot.

The Two Failures of Baseline RAG

  1. Connecting disparate information — where the answer requires joining facts across documents that each score poorly alone.
  2. Holistic semantic understanding — questions about themes across a whole collection, which no single passage answers.

Method

An LLM builds a knowledge graph from the private corpus, then “bottom-up clustering that organizes the data hierarchically into semantic clusters” produces pre-summarized themes. Retrieval traverses that structure instead of ranking isolated chunks.

Evaluation

  • Dataset: VIINA (Violent Incident Information from News Articles) — thousands of Russian and Ukrainian news articles from June 2023
  • Graph construction: GPT-4 Turbo
  • Baseline: LangChain’s Q&A
  • Graph ML: graspologic

GraphRAG “consistently outperforms baseline RAG” on comprehensiveness, human enfranchisement and diversity, holding similar faithfulness under SelfCheckGPT evaluation.

Applied domains named: social media, news articles, workplace productivity, chemistry.