Introduction to GraphRAG for Local and Global Query-Focused Summarization

GraphRAG for MSP Ticket Triage: Full Implementation Walkthrough In this video, I walk through our production implementation of GraphRAG (based on Microsoft Research's "From Local to Global" paper, Edge et al. 2024), built on a corpus of 2,523 closed support tickets and show why this architecture is central to how MSPs can turn ticket history into institutional memory. You'll see the complete pipeline in action: ▸ How closed tickets are chunked and mined into a knowledge graph of ~3,950 entities and relationships deterministically, with versioned and fingerprinted index snapshots ▸ Community detection and LLM-written community reports: how the system builds a "table of contents" for your ticket history before anyone asks a question ▸ Local queries — entity-graph retrieval for specific, cited answers your L1 team needs ("Have we seen this before?") ▸ Global queries — corpus-wide sensemaking for service delivery leaders ("What's driving our ticket volume?") ▸ Evidence gating: why the system abstains with citations instead of hallucinating when the evidence is thin ▸ The evaluation harness: anonymized pairwise LLM-as-judge (on a separate provider from the synthesizer), both presentation orders, Wilson confidence intervals, and pre-registered acceptance gates, including the honest result where we missed one gate, and what we're doing about it