HDFC Banking Intelligence Assistant (Multi-Agent GenAI System)

Overview

This project is a multi-agent AI system that intelligently routes banking-related queries between:

• Policy Q&A (RAG) — for unstructured document-based queries
• SQL Agent — for structured data queries

The system ensures accurate, grounded, and explainable responses using a combination of retrieval and agent orchestration.

---

Problem Statement

Banking users often ask two types of questions:

1. Policy-based questions (e.g., minimum balance, grievance process)
2. Data-driven questions (e.g., transaction summaries, customer stats)

Traditional systems fail because:

• RAG alone cannot handle structured queries
• SQL systems cannot interpret policy documents

This project solves that by combining both using a smart routing layer

---

Architecture

• LangGraph orchestrator
• RAG (FAISS + MiniLM)
• SQL Agent
• Streamlit UI

Features

• Intelligent query routing (RAG vs SQL)
• Grounded responses (no hallucinated answers)
• Clickable source attribution (real documents only)
• Explanation layer ("Why this answer?")
• Deployed live on Hugging Face Spaces

---

Tech Stack

• LangChain
• LangGraph
• FAISS
• HuggingFace Embeddings (MiniLM)
• Groq LLM (LLaMA 3.1)
• Streamlit

---

Key Design Decisions

• MMR Retrieval
  Improves diversity of retrieved chunks vs simple similarity search

• Source Attribution Separation
  Sources are extracted from metadata, not generated by LLM → avoids hallucinated citations

• Rule-based Routing
  Lightweight and fast; can be replaced with LLM-based router in production

---

Demo

https://huggingface.co/spaces/MLbySush/banking-intelligence-assistant

---

Future Improvements

• Replace keyword routing with LLM-based intent classification
• Add document chunk highlighting in UI
• Add evaluation metrics for RAG performance

---

Author

Sushmita — Aspiring Gen AI Engineer