VectraBank: Agentic RAG for Banking

A multi-agent AI system for comprehensive banking analysis, built with Microsoft Semantic Kernel and Retrieval Augmented Generation (RAG). The system orchestrates six specialized agents to process customer queries across fraud detection, loan eligibility, customer support, risk assessment, and strategic planning.

---

System Architecture

The system follows a layered architecture connecting user queries through a central orchestrator to specialized agents, data services, and AI models.

Key layers:

• Application Core -- Main orchestrator (main_starter.py) manages the workflow
• Specialized Agents -- Six domain-expert agents powered by Azure AI Foundry (GPT-4.1)
• Data Access Layer -- SQL Connector, Blob Connector, and ChromaDB Manager
• External Services -- Azure SQL Database, Azure Blob Storage, ChromaDB Vector Store, Azure AI Foundry

---

Agent Orchestration

Agents execute in a sequential pipeline using Semantic Kernel's SequentialOrchestration. Each agent builds on the previous agent's output, culminating in a synthesized executive report.

#	Agent	Role
1	Data Gatherer	Customer profiling, financial metrics, policy matching
2	Fraud Analyst	Transaction pattern analysis, suspicious activity detection
3	Loan Analyst	Credit risk evaluation, eligibility determination
4	Support Specialist	Customer experience assessment, retention strategies
5	Risk Analyst	Multi-dimensional risk scoring, compliance verification
6	Synthesis Coordinator	Executive report generation, strategic recommendations

---

Data Flow

The sequence diagram below shows the complete data flow from user query to final report, including interactions with Azure SQL, ChromaDB, and Azure AI Foundry.

Three phases:

1. Initialization -- Load customer profile from Azure SQL, fetch transaction data
2. Analysis -- Each agent queries ChromaDB for relevant policies and calls Azure AI Foundry for reasoning
3. Synthesis -- Coordinator integrates all findings into a structured EnhancedBankingReport

---

Project Structure

VectraBank-AgenticRAG-for-Banking/
|-- main_starter.py          # Main orchestration engine + CLI entry point
|-- blob_connector.py        # Document storage (local simulation of Azure Blob)
|-- chroma_manager.py        # ChromaDB vector database manager (6 collections)
|-- rag_utils.py             # PDF/DOCX/TXT reading, chunking, policy extraction
|-- shared_state.py          # Thread-safe state management between agents
|-- create.sql               # Azure SQL table definitions
|-- insert.sql               # Sample customer and transaction data
|-- query.sql                # Example queries
|-- pyproject.toml           # Dependencies and project metadata
|-- .env                     # Azure credentials (not committed)
|-- REFLECTION_REPORT.md     # Architecture reflection and challenges
|-- static/
|   |-- architecture.png     # System architecture diagram
|   |-- agentorchesstration.png  # Agent orchestration flow
|   |-- dataflow.png         # Sequence diagram

---

Tech Stack

Component	Technology
Agent Framework	Microsoft Semantic Kernel (SequentialOrchestration)
LLM	Azure AI Foundry -- GPT-4.1
Embeddings	Azure text-embedding-3-small
Vector Database	ChromaDB (6 banking-specific collections)
Relational Database	Azure SQL Database
Document Storage	Local file system (Azure Blob simulation)
Data Validation	Pydantic with field validators
Language	Python 3.11+

---

Prerequisites

• Python 3.11 or higher
• Azure account with:
  • Azure AI Foundry deployment (GPT-4.1 or similar)
  • Azure SQL Database with transactions table
• macOS: brew install unixodbc (required by pyodbc)

---

Setup

1. Clone and install dependencies

git clone https://github.com/Bhardwaj-Saurabh/VectraBank-AgenticRAG-for-Banking.git
cd VectraBank-AgenticRAG-for-Banking
uv venv
uv source .venv/bin/activate
uv add -r requirements.txt

2. Configure environment variables

Create a .env file with:

AZURE_TEXTGENERATOR_DEPLOYMENT_NAME=gpt-4.1
AZURE_TEXTGENERATOR_DEPLOYMENT_ENDPOINT=https://your-endpoint.cognitiveservices.azure.com/
AZURE_TEXTGENERATOR_DEPLOYMENT_KEY=your-api-key
AZURE_EMBEDDING_MODEL=text-embedding-3-small
AZURE_SQL_CONNECTION_STRING=your-sql-connection-string

3. Set up the database (optional)

Run create.sql and insert.sql against your Azure SQL Database. The system falls back to sample data if SQL is unavailable.

---

Usage

# Run everything: component tests + all 3 test scenarios
python main_starter.py --all

# Run a single demo scenario
python main_starter.py --demo

# Run all test scenarios with validation
python main_starter.py --test

---

Test Results

The system includes 29 component-level tests and 3 end-to-end scenario validations:

COMPONENT TESTS: 29/29 passed

VALIDATION SUMMARY:
  [PASS] Customer 12345: agents=6/6, risk=0.070 (low),       time=31.2s
  [PASS] Customer 67890: agents=6/6, risk=0.470 (medium-low), time=23.1s
  [PASS] Customer 11111: agents=6/6, risk=0.750 (high),       time=46.7s

Overall: ALL PASSED

Test scenarios:

Customer	Query Type	Risk Profile
12345	Financial planning & investments	Low risk -- $75K income, 780 credit score, 5 products
67890	Home loan eligibility	Medium risk -- $45K income, 680 credit score
11111	Suspicious account activity	High risk -- $28K income, 620 credit score, 1 product

---

RAG Pipeline

1. Document Ingestion -- BlobStorageConnector loads 5 banking policy documents (fraud, loans, support, risk, transactions)
2. Text Extraction -- rag_utils.read_document_file() reads PDF, DOCX, and Markdown formats
3. Chunking -- Documents are split into overlapping chunks at paragraph boundaries
4. Embedding & Storage -- Chunks are embedded and stored in 6 ChromaDB collections
5. Hybrid Search -- Combines semantic similarity with keyword boosting for retrieval
6. Context Injection -- Retrieved policies are injected into agent prompts for grounded reasoning

---

Risk Scoring

The _calculate_enhanced_risk_score() method evaluates five dimensions:

Factor	Low Risk	High Risk
Income	>= $100K (-0.15)	< $30K (+0.10)
Credit Score	>= 750 (-0.15)	< 650 (+0.15)
Customer Tenure	>= 5 years (-0.10)	< 1 year (+0.08)
Product Diversification	>= 4 products (-0.08)	<= 1 product (+0.05)
Transaction Patterns	Normal	Max > $10K (+0.10)

Score is normalized to [0.0, 1.0] and mapped to tiers: low, medium-low, medium, high, critical.

---

Sample Output

FINAL REPORT: enhanced_a1b2c3d4
Customer: 12345
Risk Assessment: low (score: 0.070)

Key Findings:
  - Customer qualifies for Tier A+ or A lending products (income: $75,000.00)
  - Excellent credit score (780) - eligible for best rates (3.5% APR)
  - High product engagement (5 products) indicates strong customer relationship

Recommendations:
  - Continue standard monitoring with annual reviews
  - Schedule periodic financial health review to identify emerging opportunities

Policy References: fraud_detection_policy_v2.md, loan_eligibility_framework.md, ...
Agent Contributions: [Data_Gatherer, Fraud_Analyst, Loan_Analyst, Support_Specialist, Risk_Analyst, Synthesis_Coordinator]

---

Documentation

• REFLECTION_REPORT.md -- Architecture decisions, challenges, and improvement suggestions
• Inline docstrings throughout all Python modules
• Structured logging to logs/banking_analysis_<timestamp>.log