〽️ UMD Student Assistant

An AI-powered chatbot that answers questions about the University of Michigan-Dearborn using a RAG (Retrieval-Augmented Generation) pipeline built with LangChain. https://umd-assistant.streamlit.app/

What it does

Students can ask natural language questions about UM-Dearborn: admissions, academics, campus life, financial aid, campus safety, and more and get accurate answers pulled directly from the university's website.

How it works

• Scraper crawls 300+ pages from umdearborn.edu using Playwright to bypass bot detection, extracts clean text, and saves each page as a .txt file
• Ingest pipeline splits each page into 500-character chunks with 50-character overlap, converts them to vector embeddings using FastEmbed, and stores all chunks in ChromaDB via LangChain
• Query takes the student's question, converts it to an embedding, and searches ChromaDB for the 3 most semantically similar chunks
• Claude receives those chunks as context via a LangChain LCEL chain and generates a natural language answer grounded strictly in scraped university data
• Memory maintains conversation context across follow-up questions using LangChain's ChatMessageHistory

Tech stack

• Python — core language
• Playwright — browser automation for scraping past Cloudflare bot detection
• BeautifulSoup — HTML parsing and text extraction
• LangChain — RAG pipeline orchestration (LCEL chains, prompt templates, retrievers, memory)
• FastEmbed — lightweight embedding model (all-MiniLM-L6-v2 via ONNX)
• ChromaDB — vector database for semantic search
• Anthropic Claude API — LLM for generating answers (claude-sonnet-4-20250514)
• Streamlit — frontend UI

Setup

git clone https://github.com/adamh36/umd-assistant.git
cd umd-assistant
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
playwright install chromium
echo "ANTHROPIC_API_KEY=your-key-here" > .env
python scraper.py
python ingest.py
streamlit run app.py

Project structure

umd-assistant/
  scraper.py      → crawls 300+ pages from umdearborn.edu using Playwright
  ingest.py       → chunks text, embeds with FastEmbed, stores in ChromaDB via LangChain
  rag.py          → LangChain LCEL chain with prompt template, memory, and generate_response()
  app.py          → Streamlit UI, calls generate_response() from rag.py
  data/           → scraped text files from 300+ UMD pages
  images/         → UI assets
  chroma_db/      → persistent vector database (gitignored)
  .env            → API keys (gitignored)

Future improvements

• Re-indexing pipeline to auto-update when UMD website changes
• Pinecone migration for cloud-hosted vector storage
• LangGraph agent with multi-tool routing