STEM Tutor RAG

An AI-powered Q&A chatbot that answers questions grounded in actual course materials, not hallucinated responses.

What this is

A retrieval-augmented generation (RAG) system that lets students ask questions about CS course topics and get answers backed by real lecture notes — with cited sources.

I originally built a system like this during my graduate work at USF for STEM courses. That code lived on university infrastructure and stayed there when I graduated. This is a clean rebuild with modern tooling.

Architecture

Lecture notes (.txt files)
  ↓
[Chunking: 500 chars, 50 overlap, sentence boundaries]
  ↓
[Embedding: all-MiniLM-L6-v2 → 384-dim vectors]
  ↓
[ChromaDB persistent store]


Student question
  ↓
[Same embedding model → query vector]
  ↓
[ChromaDB similarity search → top 3 chunks]
  ↓
[Relevance check — refuse if no good match]
  ↓
[Claude with retrieved context as the system prompt]
  ↓
Grounded answer + cited sources

Tech stack

Layer	Choice	Why
Backend	FastAPI	Async, modern, automatic docs
Embeddings	sentence-transformers (all-MiniLM-L6-v2)	Runs locally, no API cost, fast
Vector DB	ChromaDB	Easy local persistence, good for portfolio scale
LLM	Claude 3.5 Sonnet via Anthropic API	Strong instruction following, refuses to hallucinate

Key design decisions

Chunk size 500 chars with 50 overlap, breaking at sentences. Tradeoff: small enough for precise retrieval, large enough to carry context. Overlap prevents losing concepts that span a boundary. Sentence boundaries preserve semantic coherence.

Relevance threshold. If the best matching chunk is too far from the query, the system refuses to answer rather than fabricate. This is the core hallucination defense.

Source citations returned with every answer. Students can verify the answer against the original material. Trust is critical in an educational context.

Setup

# Clone and enter the project
git clone <your-repo-url>
cd stem-tutor-rag

# Create virtual environment
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

# Set your Anthropic API key
export ANTHROPIC_API_KEY=sk-ant-...

# Run ingestion (one time)
python ingest.py

# Start the server
uvicorn app:app --reload

Then open http://localhost:8000/docs to test in your browser.

Example query

curl -X POST http://localhost:8000/ask \
  -H "Content-Type: application/json" \
  -d '{"question": "What is the difference between a list and a tuple in Python?"}'

What I learned rebuilding this

• The embedding model matters more than the LLM for retrieval quality. Switching from a smaller to a more capable embedding model improved retrieval far more than swapping LLMs.
• Sentence-boundary chunking beat fixed-size chunking in my testing — the chunks contained more complete thoughts.
• The relevance threshold needs tuning per dataset. Too tight and the bot refuses good questions; too loose and it hallucinates.

Future improvements

• Hybrid search (BM25 + vector) for better recall on keyword-heavy queries
• Reranking step using a cross-encoder
• Query rewriting to handle vague student questions
• Streaming responses for better UX