AI Merchant Intelligence (Hybrid Structured + RAG)

Demo

A production-style fintech analytics assistant that answers merchant risk questions using:

1. Structured analytics computed from transaction data (ground truth)
2. RAG knowledge base (policies & explanations)
3. LLM routing to decide which system should answer each question The system answers merchant risk questions while ensuring that all financial metrics remain deterministic and auditable.

Why this project

Key idea: In real fintech risk systems, core numbers must be deterministic and auditable.
The LLM is used for explanation and context, not for inventing metrics.

Features

Structured merchant KPI computation from CSV

• total_transactions, total_revenue, average_transaction
• refunds + refund_ratio
• chargebacks + chargeback_rate
• frauds + fraud_rate
• cross_border_ratio

Policy thresholds (from knowledge base)

• refund_ratio > 30%
• chargeback_rate > 2%
• large_transactions > 10,000
• high_cross_border > 25%

RAG pipeline using LlamaIndex + HuggingFace embeddings

• FastAPI backend
• LLM-based Hybrid Router
• Streamlit UI
• Synthetic dataset generator (~1M transactions over 5 years)

Architecture (high level)

User -> Streamlit UI -> FastAPI (/ask) -> HybridRouter

• GLOBAL_STRUCTURED: dataset-wide aggregation
• MERCHANT_STRUCTURED: merchant-specific KPIs
• RAG_ONLY: definitions/policy explanations
• HYBRID: KPIs + policy explanation

Quickstart

Create environment

conda create -n rag-project-env python=3.12 -y
conda activate rag-project-env
pip install -r requirements.txt

Environment Variables

Create a .env file in the project root and add:

GROQ_API_KEY=your_groq_api_key_here

The application will not start without a valid API key.

Run the Backend API

uvicorn api:app --reload --port 8001

API will be available at:

http://127.0.0.1:8001

Swagger Docs:

http://127.0.0.1:8001/docs

Run the Streamlit UI

streamlit run app.py

Example questions

Which merchant has the highest chargeback rate?

Is merchant m_00062 above the 2% chargeback threshold?

Explain why high cross-border ratio increases fraud risk.

Compare fraud rate vs chargeback rate.

Does merchant M122 violate policy thresholds?

🚀 Quick Start

conda create -n rag-project-env python=3.12 -y
conda activate rag-project-env
pip install -r requirements.txt