Aura Agent Hackathon Submission
Thread title:
GraphImmune Trust Agent
Agent Name
GraphImmune Trust Agent
What it does
GraphImmune is a trust layer for knowledge graphs, it audits a graph before an AI/RAG system relies on it. You ask one plain question . "Is this graph safe for RAG?" and the agent reasons over a live Neo4j Aura graph, deciding which tools to run.
It finds the defects that quietly break RAG:
- Contradictions : the same subject→object asserted two opposite ways (a trusted source says "RAG reduces hallucination", a stale low-trust source says "RAG increases hallucination").
- Duplicate entities (
CeliacvsCoeliac Disease), unsupported claims (no evidence), weak relations, orphan and stale claims.
For every issue it explains why it matters using the graph's own evidence, the claim, its evidence sentence, the source document and that source's reliability then returns a Knowledge Health Score and a GraphRAG Readiness rating, and proposes safe, auditable Cypher repairs (ALIAS_OF, DISPUTED_WITH) that link and flag without ever deleting the original knowledge.
Tools (all three Aura tool types):
- Cypher Template -
find_contradictions(claims about one subject/object with opposing relations) andclaims_by_issue(claims carrying a detected trust issue). - Text2Cypher - turns ad-hoc questions ("how many claims are below 0.6 confidence?") into read-only Cypher.
- Similarity Search - fuzzy/full-text retrieval of related claims — the "what would a RAG retriever surface?" view.
It doesn't just answer - it tells you whether the knowledge is trustworthy in the first place, and shows its work.
Dataset and why a graph fits
The dataset is a document-derived trust corpus: source documents tagged by reliability (trusted / unknown / low-trust), broken into evidence-backed claims, deliberately seeded with real contradictions, duplicates, stale facts and unsupported claims.
A graph fits because trust is relational and multi-hop, not tabular:
- A contradiction is literally two paths from the same subject to the same object with opposing relations, trivial in Cypher, painful in SQL.
- Every claim traces to its evidence and source:
Evidence -[:SUPPORTS]-> Claim,Document -[:HAS_CHUNK]->...-> Evidence. - Duplicates/orphans fall out of name similarity and node degree.
- Every repair is preserved as an auditable edge, so trust decisions stay explainable.
Document -[:HAS_CHUNK]-> Chunk -[:CONTAINS]-> Evidence -[:SUPPORTS]-> Claim
Entity -[:IS_SUBJECT_OF]-> Claim -[:HAS_OBJECT]-> Entity
Claim -[:HAS_ISSUE]-> Issue -[:SUGGESTS]-> Repair
Screenshot — agent in the Aura console
Screenshot / demo - agent in action
Screenshot / demo from the website
Links
- Live app: https://graphimmune.streamlit.app
- Code: GitHub - its-serah/GraphImmune-Aura-Agent-Hackathon · GitHub