GraphRAG Concepts

GraphRAG (Knowledge Graph-based Retrieval-Augmented Generation) is an advanced RAG pattern that uses LLMs to extract a structured Knowledge Graph from unstructured text. This allows for reasoning across an entire corpus rather than just retrieving isolated snippets.

The Core Innovation: Beyond Similarity

Traditional RAG (Vector Search) is "near-sighted"; it finds similar text chunks but cannot "connect the dots" between pages that don't share semantic vectors. GraphRAG builds a Map of Knowledge where nodes are Entities and edges are Relationships.

Key Techniques for Agentic Wikis

1. Hierarchical Community Detection

Using algorithms like Leiden, GraphRAG clusters related notes into "communities."

• Application: This is the machine equivalent of a Map of Content (MOC). While humans build MOCs for navigation, an agent can use Community Summaries to understand the "Big Picture" of a vault without reading every file.

2. Global vs. Local Search

• Local Search: "Tell me about Entity X." (Explores immediate links).
• Global Search: "What are the major themes in this vault?" (Uses community summaries to synthesize a holistic answer).

3. Entity & Relationship Extraction

GraphRAG treats text as a source of Claims.

• The Seam: This is where the boundary between Flat Text and Structured Relational data dissolves. A YANP note is the source, but the extracted Graph is the "Active Memory."

Leverage for Humans-in-the-Loop

GraphRAG makes the agent's internal "world model" transparent:

• Inspectability: Humans can see the graph and correct a "broken link" or a "hallucinated relationship."
• Traceability: Every claim in the graph points back to a source TextUnit (the original note).

Comparison: Wikilinks vs. GraphRAG

References

• memory-spectrum
• llm-wiki-pattern
• agent-knowledge-vault
• hybrid-retrieval-spec