Email Classification and Response System

Project Overview

This project involves the development of an AI-driven system designed to categorize and respond to university emails. It leverages fine-tuned language models (LLMs) to classify incoming emails into three main categories:

• Student Queries
• Academic Collaboration Requests
• Corporate Inquiries

Depending on the classification, the system either generates an automatic response using Retrieval-Augmented Generation (RAG) or escalates the email to the Head of Department (HOD) for further review and manual response.

Key Features

1. Email Classification using Fine-tuned LLMs (OPT-350M):

   • Emails are classified into the three predefined categories using a fine-tuned version of Facebook's OPT model.

2. Stacked LSTM Neural Network for Classification:

   • An additional Stacked LSTM model is trained to classify emails, which allows for accurate categorization based on email body content.

   Why Stacked LSTMs + Dense Layers for Email Classification?

   a. Stacked LSTMs for Sequential Data:

   • Emails are sequences: LSTMs handle the sequential nature of emails, capturing word dependencies and context.
   • Long-range dependencies: Stacking LSTM layers helps learn both low-level (word) and high-level (sentence) patterns, improving understanding.

   b. Dense Layers for Classification:

   • Feature abstraction: After LSTMs, dense layers process the learned features and map them to categories.
   • Non-linearity: Dense layers help form complex decision boundaries for better classification accuracy.

   c. Handles Email Complexity:

   • Varied content: LSTMs adapt well to different email lengths and tones.
   • Context understanding: LSTMs learn relationships across the text, while dense layers ensure accurate predictions.

3. Retrieval-Augmented Generation (RAG):

   • RAG is used with LLama-3.2-1B for generating automated responses to student and academic inquiries.

4. Document Search with FAISS:

   • A FAISS-based similarity search helps in retrieving relevant information from large documents, enabling the system to provide accurate and relevant automated responses.
   • The automated responses are divided into two types: - Actual response from the RAG using LLMs for student inquiries and academic enquiries if they are present in the database of the university. - Hardcoded response for Corporate level Inquiries.

5. Interactive Gradio UI:

   • A Gradio-based interface is provided for users to input email queries, classify them, and generate automated responses.

---

Installation

Prerequisites

Ensure that you have Python installed on your machine. You can install the required dependencies by running the following command:

pip install datasets sentence_transformers PyMuPDF PDFReader pdfplumber faiss-cpu --no-cache langchain pypdf langchain-community streamlit huggingface_hub gradio -U

How to Run the Application

1. Pull the Docker Image from Docker Hub

The Docker image anishkarnik/smartsense_ta:v1 is already built and pushed to Docker Hub. You can pull the image directly using the following command:

Note: I have pushed the image on my friend's docker because my docker had some issues

bash
docker pull anishkarnik/smartsense_ta:v1

docker run -e HUGGINGFACEHUB_API_TOKEN=<your-token> -p 7860:7860 
anishkarnik/smartsense_ta:v1

http://localhost:7860