Every developer building GraphRAG systems will face a hard question: which graph should an agent trust when different models, prompts, and ontologies extract different entities from the same data?
In enterprise and financial settings, this creates duplicated entities, fragmented context, and conflicting facts across teams, tools, and model providers.
In this session, the speaker will share a real experiment on ontology-guided graph federation for agentic RAG. You will learn:
1. How model, prompt, and ontology choices change the entities and facts extracted from the same source documents.
2. Why entity duplication is not just a cleanup problem, but a signal that different business contexts may need to be preserved.
3. How Neo4j multi-database federation and provider-aware routing can help agents compare, select, and combine evidence without flattening provenance.
The speaker will walk through lessons from scaling a financial extraction benchmark from 16 to 80 cases across three indexing profiles and four LLM providers.
In the larger run, two ontology/prompt profiles produced over 1,200 extracted facts each, compared with 305 in the baseline, while preserving provider-level lineage for later MDM and survivorship analysis.
You will leave with a practical mental model for using graph databases in the GraphRAG era: not as one final memory store, but as a federated context layer where agents can reason across competing views of the same data.
Speaker
Ii tae Jeong
XCENA
Ii Tae Jeong is a Graph Product Engineer specializing in graph query engines with hardware acceleration. With a strong focus on scalable graph systems, he has delivered over four tutorial and workshop sessions at academic conferences for both researchers and practitioners. His work bridges advanced graph technologies with real-world applications, aiming to push the boundaries of performance and usability in modern data systems.
