Project 42
Start learning
Module 3 of 5 · 55 min

GraphRAG & Multi-Hop Entity Reasoning

Extract structured knowledge graphs from unstructured corpora to enable multi-hop reasoning and holistic summary queries across disparate documents.

Core concept

By the end

You will be able to

  • Explain why standard vector RAG fails on global summarization and multi-hop entity relationships.
  • Extract entities, relationships, and claim triples using structured LLM extraction.
  • Index and cluster knowledge graphs using hierarchical community detection algorithms (Leiden).
  • Execute GraphRAG query workflows combining local entity neighborhoods with global community summaries.
01

The Multi-Hop and Global Query Blind Spot

Standard vector RAG searches for localized text snippets matching a specific semantic vector. It fails on queries like "What are the top 5 emerging risks across all 200 audit reports?" because the answer is distributed across thousands of disjoint chunks.

GraphRAG extracts an interconnected knowledge graph of entities, relationships, and hierarchical communities, synthesizing global summaries at each community level.

Extracted Entity-Relation Triple Schema
json
{
  "entities": [
    { "name": "Project 42", "type": "ORGANIZATION", "description": "Open source AI education platform" },
    { "name": "Orchard", "type": "SYSTEM", "description": "Autonomous discovery and maintenance pipeline" }
  ],
  "relationships": [
    { "source": "Orchard", "target": "Project 42", "relationship": "MAINTAINS_CURRICULUM_FOR", "strength": 0.95 }
  ]
}

Practice activity

Extract Entity-Relationship Graph from Unstructured Text

  1. Process a 5-document incident corpus through an entity extraction prompt.
  2. Construct an adjacency matrix and query multi-hop relationships.
  3. Generate a cross-incident root cause summary using community reports.

What to produce

  • Graph JSON export containing nodes, edges, and community cluster descriptions.

Reflect before continuing

What are the cost and latency trade-offs of building a knowledge graph upfront during ingestion versus querying vector databases on demand?

Evidence

Sources and verification

Knowledge check

Make it stick.

Pass at 80%

Choose the strongest answer for each question. Your attempts become part of your account transcript.

01What type of query benefits most from GraphRAG community summaries over basic vector RAG?