Alfred RAG Assistant

Useful student information is scattered across documents, while general chatbots can answer confidently without evidence. I wanted a narrow assistant whose answers could be traced back to known sources.

I built a Python RAG pipeline over a 10-document knowledge base, producing 47 embedded segments in ChromaDB and retrieving the top four sources for Groq's Llama 3.3 model. The system includes source attribution and strict refusal behavior.

• Evaluated at 80% response accuracy on a focused question set.
• Added grounded citations and refusal behavior to reduce hallucinated answers.