Problem Statement: University websites contain vast amounts of information (curricula, ECTS guidelines, opening hours), but finding specific answers is often difficult due to fragmented navigation. Standard keyword searches fail to answer complex questions, such as "Which courses do I need for the Data Science specialization?"
Project Goal: The goal is to develop a Retrieval-Augmented Generation (RAG) Chatbot. Unlike basic scripted bots, this system will ingest official university PDF/HTML documents, index them, and use an LLM to generate accurate, cited answers to student queries in natural language.
Core Tasks:
- Knowledge Ingestion: Build a scraper/parser to extract text from Curriculum PDFs and University Webpages and chunk them for indexing.
- Vector Search Pipeline: Implement a retrieval system (using LangChain and a Vector DB) that finds the most relevant document chunks for a given student question.
- Hallucination Guardrails: Engineer the LLM system prompt to strictly answer only based on the retrieved context and provide citations (e.g., "Source: Curriculum 2024, Page 5").
- Chat Interface: Build a user-friendly web widget (React) that maintains conversation history and context.
Research Focus:
The research focus lies on an Answer Quality Evaluation. The student must curate a dataset of ~30 "Golden Questions" and measure the bot's performance using metrics such as Answer Relevance and Faithfulness (e.g., using frameworks like RAGAS or manual expert review).
Technologies: Python (LangChain); Vector Databases (ChromaDB); LLMs (OpenAI/Ollama); React
Tags: Natural Language Processing; Information Retrieval; RAG; Chatbots; EdTech; Web Engineering; Kalchgruber
