All case studies / Case study 03 / 12
Veltaro Industries
GraphRAG knowledge engine
A knowledge engine that connects Jira, Confluence and Teams
In short
Years of tickets, wiki pages and conversations turned into a knowledge graph that teams query in plain words, with answers that connect the information together and cite their sources.
- Client
- Veltaro Industries
- Sector
- Manufacturing, automotive parts
- Size
- 1,200 staff, 3 plants
- −70%time spent looking for information
- 3 toolsconnected in a single answer
- 100%of answers cite their sources
The context
Veltaro’s quality, methods and maintenance teams document everything: incidents in Jira, procedures in Confluence, decisions in Teams conversations. But this knowledge lives in silos, and the right answer often spans three tools.
The challenge
- 01
A classic search engine finds words, not cause-and-effect links.
- 02
A standard RAG assistant handles simple questions well, and questions that need several documents connected poorly.
- 03
Experts retire and take the context of past decisions with them.
- 04
The data has to stay on the company’s infrastructure.
Our approach
An LLM reads the documents and extracts entities (machines, parts, incidents, procedures, decisions) and their relationships into a Neo4j graph. For each question, a hybrid retrieval combines the graph, to follow the links, and vector search, for meaning. The model then writes an answer that explains how the facts connect and cites every source.
The solution, step by step
- 01
Connected to the sources
Jira, Confluence and Teams are synced every night, respecting everyone’s access rights.
- 02
Entities extracted
The model spots machines, parts, incidents and procedures, and how they relate.
- 03
A knowledge graph
Everything is linked in Neo4j: the incident points to the machine, the procedure and the decision taken.
- 04
Hybrid retrieval
The question follows graph links and semantic closeness to find the full chain of facts.
- 05
A connected, sourced answer
The assistant explains how the facts connect and cites each ticket, page or conversation.
Results
- −70%search time for the methods teams
- 2 weekssaved onboarding a new engineer
- 12,000documents connected in the graph
- 0data leaving the infrastructure
What we deliver
- Jira, Confluence and Teams connectors
- Entity and relationship extraction pipeline
- Neo4j graph and vector index
- FastAPI API and Q&A interface
- Answer quality dashboard
Technologies
- Neo4j
- LlamaIndex
- Python
- Hugging Face embeddings
- FastAPI
- Docker
Timeline
10 to 14 weeks
Team
2 AI engineers, 1 data engineer
Illustrative cases: clients and figures are fictional, representative of the solutions we deliver.