🧠 RAG-based Loan Approval Q&A Chatbot

This project implements a lightweight Retrieval-Augmented Generation (RAG) chatbot designed to answer questions related to a Loan Approval Prediction Dataset. It combines vector-based semantic search with generative language modeling using a TinyLlama model for fast, contextual, and informative responses.

🚀 Project Overview

The chatbot enables users to ask natural language questions about a tabular dataset, returning informed answers based on both statistics and sample data. It uses:

• SentenceTransformers for embedding dataset chunks.
• FAISS for vector similarity search.
• TinyLlama-1.1B-Chat for generating responses.
• Gradio for building a user-friendly chat interface.

📁 Dataset Used

The notebook expects a CSV file named Training Dataset.csv, representing a typical loan application dataset with features like:

• Income
• Credit History
• Employment Status
• And others...

Each column is described statistically using pandas.describe() and enriched with a few sample rows.

🧱 Tech Stack

Component	Description
pandas	Data loading and processing
sentence-transformers	Converts text chunks to embeddings
faiss-cpu	Efficient similarity search over embeddings
transformers	Loads the TinyLlama text generation model
gradio	Provides a web-based chat UI

🔍 How It Works

1. Data Chunking: Each column's stats and a few sample rows are turned into textual "chunks" to form a knowledge base.
2. Vector Indexing: These chunks are embedded and stored using FAISS for nearest-neighbor search.
3. Query Retrieval: At runtime, the chatbot retrieves top-K relevant chunks based on semantic similarity to the query.
4. Contextual Answering: These chunks, along with chat history, are fed into a TinyLlama-based text generation pipeline to produce an answer.

Snippet

🧠 Prompt Template

The prompt to the model includes:

• Last 3 turns of conversation (user + bot)
• Top-K retrieved dataset chunks
• A task-specific instruction:

You are an assistant for questions about a loan approval dataset.
Previous conversation:
User: ...
Bot: ...
User: <query>
Based on the context below, answer the user’s question.

Context:
<retrieved chunks>

Answer:

🖥️ Launching the App

To run the chatbot interface locally with sharing enabled:

!pip install pandas sentence-transformers transformers faiss-cpu gradio

Then execute the notebook and follow the Gradio link:

gr.ChatInterface(...).launch(share=True)

✅ Sample Query

rag_answer("What features are most important for loan approval?", "")

This will produce a model-generated response grounded in dataset features like income, credit history, etc.

📌 Notes

• Ensure the dataset file is named correctly (Training Dataset.csv).
• TinyLlama is used for fast inference; this can be swapped with a larger model if needed.
• No external APIs or cloud resources are required—everything runs locally.