Danielfonseca1212/Benchmark

🕸️ GraphRAG vs Vector RAG — Live Fraud Detection Benchmark

By [Daniel Fonseca](https://linkedin.com/in/daniel-fonsecaai) · AI/ML Engineer · Graph Neural Networks · Fraud Detection

  

What this demo shows

A live benchmark comparing two RAG architectures on fraud detection queries:

Core insight: Fraud lives in connections. A device shared by 3 customers, a money mule chain with 3 hops, 6 accounts from the same IP — these patterns are invisible to embeddings but trivially discoverable with a single Cypher traversal.

Architecture

User question (natural language)
        │
        ▼
Groq/Llama 3.1 ──► Cypher query generation
        │
        ▼
Neo4j Aura ──► Graph traversal (2-5 hops)
        │
        ▼
Structured records ──► Groq/Llama ──► Fraud analysis answer

Graph schema

(Customer)-[:HAS_ACCOUNT]->(Account)
(Customer)-[:USED]->(Device)
(Account)-[:ACCESSED_FROM]->(IP)
(Account)-[:TRANSFER {amount, date}]->(Account)
(Account)-[:TRANSACTION {amount, type}]->(Merchant)

Fraud patterns detectable:

• —🔴 Shared device cluster — emulator farms, identity theft
• —🔴 IP overlap — account opening fraud
• —🔴 Money mule chain — layering (A-102 → A-445 → A-667 → A-890)
• —🔴 Card testing — micro-transactions on merchants

Setup (add to HF Secrets)

After adding secrets: click "Seed fraud graph" in the sidebar to populate Neo4j.

Without credentials the app runs in demo mode with realistic simulated responses.

Related projects

• —IBM Safer Payments — AUC-ROC 0.9591
• —HetGNN Fraud Graph Explorer
• —Agentic RAG Pipeline on Kubernetes

Built with Neo4j Aura · Groq · Llama 3.1 · Streamlit · PyVis · Plotly