Here’s a markdown documentation draft for your publication:

---

AI-Powered University Chatbot Using RAG Techniques

Abstract

This project explores the creation of an AI-powered chatbot tailored for universities, leveraging Retrieval-Augmented Generation (RAG) techniques. The chatbot integrates advanced language models and semantic search to provide accurate, contextually relevant answers to user queries. It is specifically designed to support tasks such as answering university-related FAQs, assisting with educational content, and streamlining administrative interactions.

---

Introduction

Modern educational institutions require efficient tools to manage communication and address user queries. Our chatbot combines semantic search with generative AI to provide precise answers, sourced from university-specific datasets, using a robust backend powered by PostgreSQL with pgvector.

---

System Architecture

The system consists of the following components:

1. User Query Interface: Built using Flask, enabling users to interact with the chatbot.
2. Database: PostgreSQL with the pgvector extension to store question embeddings for semantic search.
3. External APIs: Integrates with external datasets and language models for retrieval-augmented generation.

• Image source

---

Models and Datasets Used

Models

• Sentence Embedding Model:

  • Model: all-MiniLM-L6-v2 from Hugging Face.
  • Usage: Generates embeddings for similarity-based search.

• Language Model:

  • Model: llama-3.3-70b-versatile accessed via Groq API.
  • Usage: Generates responses using retrieved context.

Datasets

• Embedded Questions Dataset:
  • Source: Self-curated and stored in PostgreSQL.
  • Details: Includes questions, topic IDs, and embeddings for retrieval.

from sentence_transformers import SentenceTransformer
model = SentenceTransformer('all-MiniLM-L6-v2')
embedding = model.encode("Sample question")

---

Implementation Details

Backend

The backend is built using Flask, enabling efficient communication with the database and external APIs. It integrates the following key functionalities:

• Semantic search using pgvector.
• Feedback collection and storage for improving response accuracy.

Retrieval and Response Generation

1. Embedding Similarity Search:
   Retrieves top-n similar questions based on vector similarity.

2. Response Generation:
   Combines the retrieved context with a user query to generate responses.

SELECT sr_no, topic_id, question, 1 - (embedding <=> %s::VECTOR) AS similarity
FROM questions
ORDER BY similarity DESC
LIMIT %s;

---

• Image source

Challenges and Solutions

1. Challenge: Data Collection.

   • Solution: Worked as a team, collected data from university students, used web scrapping.

2. Challenge: Efficiently handling user feedback.

   • Solution: Developed a feedback logging system to refine response accuracy over time.

---

Evaluation Metrics

• Accuracy: Measured as the percentage of correctly retrieved results.
• User Feedback: Positive vs. negative feedback to analyze chatbot performance.

---

Conclusion

This project demonstrates how RAG techniques can enhance chatbot performance in educational settings. By integrating semantic search and generative AI, the chatbot provides contextually relevant responses, improving user satisfaction and engagement.

---

Live Link

Future Work

1. Enable storing and referencing previous responses to provide context-aware answers.
2. Develop mechanisms to handle user queries with grammatical errors or irrelevant content.
3. Implement a robust feedback-based system for continuous improvement, incorporating user likes and dislikes.
4. Add a classifier model to categorize queries into domains like academics, curriculum, placements, facilities, etc.

Suggested Links:

• GitHub Repository for source code.
• Documentation for pgvector.

---