Thesis · livefromeurope
Dynamic Knowledge Graph Integration for Enhanced News Event Summarization
BaselineRAGversusGraphRAG
Turning raw news into a living knowledge graph, so AI can explain how events connect.
Georg Johannes Talaszlivefromeurope.eu
01 · The problem
AI only knows the past
Language models are frozen at their training date. Classic RAG adds fresh articles, but one by one, and misses how events relate.
“Machine learning models that are trained on previous news articles are only able to speak to the past.”
Event correlation
How are event A and event B connected?
Temporal correlation
Which event happened first?
02 · The idea
From news to a knowledge graph
GraphRAG links events by meaning and by time. Five steps, from raw articles to an answer.
Step 1/5 · Detect events
1Detect events
Thousands of articles are embedded and clustered. Each cluster is one event. Noise is dropped.
- Embeddings
- UMAP
- HDBSCAN
- Outlier removal
2Build the graph
People, places and organisations become nodes, their relations become links. Russia appears in both elections: a hidden connection.
- Relik
- Entity linking
- Relation extraction
3Find communities
The Leiden algorithm finds tightly connected groups. An LLM writes a summary for each one.
- Leiden algorithm
- Community summaries
4Retrieve & reason
A question pulls in the relevant communities and the paths that connect them.
- Community retrieval
- LLM context ranking
5Answer
An agent writes one summary that connects events across days: how they relate, and what came first.
- LangChain agent
Live now
Today's European news as a knowledge graph
Stories, people, places and themes, and how they connect.
03 · Experiment
BaselineRAG vs GraphRAG
79%
GraphRAG wins with LLM judges
520 evaluations · GPT-4o & Mistral Large · p < 0.001
69%
GraphRAG preferred by experts
130 evaluations · 13 experts · p < 0.05
EU policy officer, journalist, legal and data professionals, praising clarity, contextual depth and relevance.
50% = no difference between the two
04 · Conclusion
Deeper answers, at a cost
Gain
More context-rich, connected summaries across days and events.
Cost
16×slower graph construction
6×slower answers
Best for
In-depth summaries and retrospective analysis, less for real time.