🌱 Sage — RCC User Guide Assistant

A chat assistant for the University of Chicago's Research Computing Center. Ask about accounts, SSH, Slurm jobs, storage, and software — answers are retrieved from the official RCC User Guide, with citations.

Built with Streamlit and the Mistral API. Read-only and RAG-only — it reads documentation, never runs commands on the cluster.

Features

• 🔎 Grounded answers — retrieves and cites real RCC docs; no hallucinated commands.
• 📎 File uploads — drop in a PDF or text file and ask about it.
• 💬 Streaming replies with conversation memory for follow-ups.
• 🗂️ Self-updating docs — a one-command refresh keeps the bundled guide current.

Quickstart

pip install -r requirements.txt
export MISTRAL_API_KEY=sk-...
streamlit run app.py            # → http://localhost:8501

Configuration

Variable	Default	Purpose
MISTRAL_API_KEY	(required)	Mistral API key
RCC_DOCS_PATH	./docs	User Guide markdown source
RCC_WEB_PATH	./web	Scraped website text source
LOG_LEVEL	WARNING	Python log level

How it works

The model is given two tools — search_docs(query) to rank the documentation tree and read_doc(path) to read a page — so every file under docs/ and web/ is reachable, and answers stay tied to the source. Reads are confined to those directories.

Keeping docs fresh

The bundled docs//web/ snapshot is refreshed from the upstream user-guide repo. Run from an internet-connected host:

./refresh-docs.sh            # sync docs + web, stamp docs_snapshot.json
./refresh-docs.sh --scrape   # also re-scrape the RCC website

Layout

app.py             # the Streamlit app (UI + RAG loop)
refresh-docs.sh    # pull upstream docs → docs/ + web/
docs/              # RCC User Guide (markdown)
web/               # scraped RCC website (text)
docs_snapshot.json # source commit + sync date