UBAssist is a retrieval-augmented conversational agent that gives accurate, source-grounded answers about the University at Buffalo, with special focus on the needs of international students (F-1/OPT, I-20, course selection, housing, etc.). It combines:
| Layer | Tech |
|---|---|
| Retriever | FAISS dense vectors (BAAI/bge-small-en-v1.5) + keyword hybrid search |
| Reranker | cross-encoder/ms-marco-MiniLM-L-6-v2 |
| Fine-tuning | LoRA / QLoRA adapters on microsoft/phi-1_5 with RAFT |
| API | Flask (backend) |
| UI | Flask (+ vanilla JS) single-page chat |
- Context-aware chat – tracks the last few turns for coherent follow-ups.
- Grounded citations – every answer lists clickable sources.
- Lightweight adapters – fine-tuned with LoRA/QLoRA; base model stays frozen.
- One-click deploy – Dockerfile +
start.shspin up scraper, vector-store, API & UI. - Scalable vector DB – FAISS index stored in
data/embeddings/.
👉 Try it live: https://huggingface.co/spaces/TeamSAS/UB_VSA
We’re on the free HF hardware tier; first response may take ~30 s while the container wakes up.
buffalo_rag/
├── api/ # FastAPI routes & schemas
├── embeddings/ # Chunker + sentence‑transformer encoder
├── frontend/ # Flask templates & static JS/CSS
├── model/ # BuffaloRAG orchestration
├── scraper/ # Playwright / BeautifulSoup crawler
└── vector_store/ # FAISS DB + hybrid search logic
data/
└── embeddings/ # .pkl embeddings + faiss_index.pkl (generated)
Dockerfile
start.sh
main.py # CLI pipeline entry‑point
git clone https://huggingface.co/spaces/TeamSAS/UB_VSA
cd UB_VSA
# 1. Python deps
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
export HUGGINGFACEHUB_API_TOKEN=hf_your_token_here
# 3. Build embeddings (only first time, ~5 min)
python main.py --build
# 4. Launch backend + frontend
python main.py --run