LLM Query Agent

A conversational AI agent capable of answering questions based on lecture notes and other resources using advanced NLP techniques and pre-trained language models. Link to the approach

Table of Contents

• Introduction
• Features
• Installation
• Usage
• Data Collection and Preprocessing
• Creating and Storing Embeddings
• Building the Query Agent
• Future Work

Introduction

This project aims to create an AI agent that can answer questions from a specific context (lecture notes, web resources). The agent utilizes NLP techniques and pre-trained language models to provide accurate and meaningful responses.

Features

• Data Collection from PDFs and websites
• Data Preprocessing using NLP techniques
• Embedding creation and storage using FAISS
• Query processing and response generation using Google's Gemini Pro LLM
• Follow-up question handling using Langchain's Conversation Buffer Memory
• User interface built with Streamlit

Installation

Prerequisites

• Python 3.9+
• Pip

Steps

# Clone the repository
git clone https://github.com/DevG10/Ema-s-NLQ.git

# Navigate to the project directory
cd Ema-s-NLQ

# Install the dependencies
pip install -r requirements.txt

Usage

Run the main.py file

cd src
python -m main.py

Note: Before running the main file ensure that you have first created a .env file and pasted your gemini key as GOOGLE_API_KEY = 'your_api_key' inside your .env file

Data Collection and Preprocessing

Data Sources

• Lecture Notes
• Websites from well-known institutes

Tools and Techniques

• BeautifulSoup for web scraping
• PyMuPDF for PDF extraction
• NLTK for preprocessing (stopwords removal, lemmatization, etc.)

Creating and Storing Embeddings

• Embeddings convert text into numerical vectors, allowing for semantic search and retrieval.
• To create and store the embeddings, I used FAISS which is open-source and fast too as it is made using C++
• The embeddings were calculated using sentence transformer's all-MiniLM-L6-v2 pre-trained model

Building the Query Agent

Query Processing

• The query given by the user is first converted to the embeddings using the same model which was previously used to create the embeddings of the lecture.
• This converted query is then searched using vector similarity search and the matching docs are returned

Handling Follow-Up Questions

• To enable follow-up questions, Langchain's Conversation Buffer Memory is used to keep track of the context and questions asked.

Future Work

• Potential improvements and additional features:
• Enhancing the accuracy of responses
• Adding more data sources
• Improving the user interface