RAG-Based Learning & Code Assistant

Abstract

The RAG-Based Learning & Code Assistant is a dual-purpose AI application designed to provide academic support to students and technical guidance to developers. Leveraging the power of LangChain, Groq's LLaMA 3, and ChromaDB, the system processes and retrieves information from user-uploaded documents. With an intuitive Gradio interface and fallback TF-IDF embeddings, the assistant ensures contextual and reliable responses. This work also evaluates retrieval performance using standard metrics such as Recall@k, Response Accuracy, and User Satisfaction Score, and compares results with alternative RAG setups to establish robustness and effectiveness.

---

Introduction

Retrieval-Augmented Generation (RAG) combines document retrieval with large language models to generate grounded and context-aware responses. This project implements a lightweight RAG-based assistant that serves two key domains:

• Education (Learning Tutor): Assisting students with personalized academic content
• Software Development (Code Helper): Assisting developers in code comprehension and documentation support

Built using LangChain, Groq's LLaMA 3, ChromaDB, and Gradio, the system is designed for accessibility, modularity, and speed.

---

Methodology

Document Processing

• Accepts .pdf, .txt, .md, .py, etc.
• Splits content into manageable chunks using LangChain’s RecursiveCharacterTextSplitter
• Generates vector embeddings using HuggingFace models (or fallback TF-IDF)
• Stores embeddings in ChromaDB with metadata (e.g., source, page)

Query Flow

• User enters a query
• System retrieves top relevant chunks (k=4)
• Constructs a context-aware prompt using retrieved chunks
• Sends the query + context to LLaMA 3 (via Groq API)
• Gradio interface returns a conversational response with citation support

Modes

• Learning Tutor: Academic materials like lecture notes, books, or assignments
• Code Helper: Programming content like Python files or documentation

---

Diagrams

• RAG Retrieval Architecture
  

• ChromaDB Retrieval Process in RAG

Evaluation Methodology

To assess retrieval and response quality, we used the following:

1. Retrieval Evaluation Metrics

• Recall@k: Measures how many ground-truth relevant chunks are in the top-k retrieved documents.
• Precision@k: Evaluates the relevancy of the top-k retrieved documents.
• Average Context Overlap: Measures semantic overlap between retrieved content and expected content.

2. LLM Response Evaluation

• Response Accuracy (Human-verified): % of responses factually consistent with source.
• Hallucination Rate: % of LLM responses with unverifiable or incorrect claims.
• User Satisfaction Score: Users rated responses on a 5-point scale for helpfulness and relevance.

3. Testing Procedure

• 10 academic files (physics, biology, CS)
• 10 codebases (Python projects, scripts)
• 5 sample queries per file
• Evaluated both HuggingFace embeddings and TF-IDF fallback

---

Results

Observations:

• LLaMA 3 produced structured, context-rich answers when paired with semantic embeddings.
• TF-IDF was lightweight and acceptable where transformers were unavailable.
• Students found explanations clear; developers appreciated inline code references.

---

Comparison with Existing RAG Systems

---

Real-World Applications

• Student Use Case: A CS student uploads OOP lecture slides and asks, “Explain abstraction with example code?” → receives a detailed explanation + code.
• Developer Use Case: A developer uploads a Python script and asks, “What is this function doing?” → gets a summary with inline comments.
• Offline Setting: Rural students used TF-IDF fallback when internet was limited.

---

Live Application on Hugging Face

Try it here: RAG-Based Learning & Code Assistant

---

Conclusion

The RAG-Based Learning & Code Assistant demonstrates how a lightweight, modular, and dual-purpose RAG system can be built using modern tools like LangChain, Groq’s LLaMA 3, and ChromaDB. It supports both academic and technical use cases with fallback mechanisms for robustness.

---

Acknowledgements

We thank the LangChain, HuggingFace, and Groq communities for their open-source contributions that enabled this work. Feedback from real users during testing helped refine our evaluation strategy and improve practical usability.

---