Banking FAQ Assistant

A production-grade Agentic AI Capstone project built with LangGraph, ChromaDB, and Streamlit. This Banking FAQ Assistant handles customer queries about account types, loans, interest rates, fund transfer charges, and more — powered by a retrieval-augmented generation (RAG) pipeline with faithfulness evaluation.

---

Project Structure

banking-faq-assistant/
├── bank_bot/
│   ├── __init__.py
│   ├── state.py
│   ├── config.py
│   ├── tools.py
│   ├── nodes.py
│   └── graph.py
├── knowledge_base/
│   ├── savings_account_types.txt
│   ├── fixed_deposit_rates.txt
│   ├── rtgs_neft_imps_charges.txt
│   ├── personal_loan_features.txt
│   ├── home_loan_eligibility.txt
│   ├── gold_loan_eligibility.txt
│   ├── credit_card_variants.txt
│   ├── current_account_variants.txt
│   ├── atm_debit_card_charges.txt
│   ├── locker_facility_charges.txt
│   ├── kyc_account_opening.txt
│   └── insurance_products.txt
├── tests/
│   └── ragas_eval.py
├── capstone_streamlit.py
├── ingest.py
├── requirements.txt
├── .env
└── README.md

---

Setup & Installation

1. Prerequisites

• Python 3.10 or higher
• A Groq API Key

---

2. Installation

Clone the repository and set up the virtual environment:

git clone https://github.com/your-username/banking-faq-assistant.git
cd banking-faq-assistant
python -m venv venv

# Windows
venv\Scripts\activate

# Mac/Linux
source venv/bin/activate

pip install -r requirements.txt

---

3. Environment Variables

Create a .env file in the root directory:

GROQ_API_KEY=your_actual_api_key_here

---

4. Knowledge Base Setup

Ensure your 12 banking documents are placed inside the /knowledge_base folder. These files cover:

• Savings Account Types (Regular, Premium, Senior, Salary, Women's, Minor)
• Loan Products (Home Loan, Personal Loan, Gold Loan)
• Fixed Deposit Rates and Charges
• Fund Transfer Services (RTGS, NEFT, IMPS, UPI)
• Credit Card Variants (Classic, Signature, Infinite)
• Current Account Variants (Basic, Business, Premium, Trade)
• ATM & Debit Card Charges and Limits
• Locker Facility Charges and Allotment Policy
• KYC Requirements and Account Opening Procedures
• Insurance Products (Life, Health, Motor, Home)

---

5. Ingest the Knowledge Base

Before running the app for the first time, populate the ChromaDB vector store:

python ingest.py

---

Usage

Run the Streamlit application:

streamlit run capstone_streamlit.py

The app will open in your browser at http://localhost:8501.

---

Example Test Queries

Category	Query	Expected Behaviour
Persistence	"I live in a rural area." → "What is my MAB?"	Returns ₹2,000 (rural MAB for Regular Savings)
Logic	"What is the interest rate for a self-employed architect with a CIBIL score of 780 for a personal loan?"	Returns 13.00% (750–799 band + 0.50% self-employed premium)
Tool Use	"Calculate EMI for ₹10 lakh at 9.5% for 5 years"	Runs loan_calculator tool and returns monthly EMI
Date/Time	"What is today's date?"	Invokes get_current_datetime tool
Safety / Out-of-Scope	"How do I book a movie ticket?"	Returns helpline fallback (1800-BANK-HELP)
Greeting	"Hello!"	Friendly greeting without KB lookup

---

Agent Flow

The LangGraph StateGraph routes every query through exactly 8 nodes:

memory_node → router_node → retrieval_node → answer_node → eval_node → save_node
                          ↘ tool_node ↗
                          ↘ skip_node ↗

Node	Responsibility
memory_node	Sliding-window history + customer name extraction
router_node	Routes to retrieve, tool, or skip
retrieval_node	ChromaDB top-K semantic search
skip_node	Handles greetings and trivial inputs
tool_node	Runs loan_calculator or get_current_datetime
answer_node	Generates grounded answer from retrieved context
eval_node	Faithfulness check with retry logic (max 2 retries)
save_node	Terminal node; extensible for audit logging

---

RAGAS Evaluation

Run the baseline evaluation against 10 curated test cases:

# Full RAGAS evaluation (requires OpenAI API key and embeddings)
python tests/ragas_eval.py

# Quick smoke test (no embeddings cost, ideal for CI)
python tests/ragas_eval.py --smoke

Metrics evaluated:

Metric	Description
Faithfulness	Is the answer grounded in the retrieved context?
Answer Relevancy	Does the answer address the question asked?
Context Precision	Are the retrieved chunks relevant to the question?
Context Recall	Does the retrieved context cover the ground truth?

Results are saved to tests/ragas_results.json after each run.

---

Design Principles

• Strict grounding — The answer_node system prompt forbids introducing facts absent from the retrieved context.
• Faithfulness loop — eval_node re-runs answer_node up to MAX_EVAL_RETRIES times if hallucination is detected.
• Resilient tools — loan_calculator and get_current_datetime never raise exceptions; they return error strings.
• Lazy singletons — LLM client and ChromaDB collection are initialised once via @lru_cache and @st.cache_resource.
• Sliding memory — Only the most recent MEMORY_WINDOW (default: 6) turn-pairs are sent to the LLM to manage token cost.

---

License

This project is developed for educational purposes as part of a Senior AI Engineering Capstone. Not intended for production banking use.