On-Premise GraphRAG Incident Knowledge Base
A GraphRAG system, running fully on-premise via Docker with no dependency on external APIs, that responds to natural-language questions with the closest matching solutions and points to check.
Duration
Contract development (2 weeks)
Team
1-2 people
Client
A major infrastructure group
01 // OVERVIEW
Project Overview
A prototype GraphRAG incident-knowledge system developed for an infrastructure company group. It extracts causal relationships (cause -> action -> result -> countermeasure) from failure-case data and stores them in a Neo4j graph, then presents the closest matching solutions and points to check in response to natural-language queries. The requirement was fully on-premise operation via Docker Compose, with no dependency on external cloud APIs.
02 // KEY FEATURES
Key Features
Building and Editing the Causal Relationship Graph
Automatically extracts causal relationships from case data and stores them in Neo4j. Graph-editing features - adding, deleting, and reweighting nodes and edges, and merging similar concepts - allow the knowledge base to be maintained continuously.
Natural-Language Incident Q&A
Entering a natural-language question such as "What should I check if incident X occurs?" returns the related solutions and checkpoints from the graph. Uses a Llama 3-family model running on Ollama on-premise.
Fully On-Premise, Complete via Docker Compose
Designed so all functionality works without an internet connection. All services launch with a single Docker Compose command, making the setup portable directly into the customer's server environment.
03 // TECHNICAL DEEP-DIVE
Technical Highlights
Solution
Operating an LLM Under a Zero-External-API Constraint
Since cloud APIs such as OpenAI could not be used, we ran Ollama + Llama 3 8B (4-bit quantized) locally on an NVIDIA GPU (A4000, 16GB VRAM), achieving practical inference speed in an on-premise environment.
Solution
Incorporating Terminology Variation into the Graph
Implemented a feature for merging identical and similar nodes so that expressions deviating from the public database's standardized terminology could still be incorporated. Designed so that slang and abbreviations used on-site correctly reach the right related nodes.
04 // IMPACT & RESULTS
Results
2 weeks
Prototype delivery time
No external APIs
Fully on-premise operation achieved
Docker-based
All services launch with docker compose up