README.md

metadata

title: StudyHub
emoji: 📚
colorFrom: indigo
colorTo: purple
sdk: gradio
sdk_version: 5.50.0
app_file: app.py
pinned: false

StudyHub

A private, multilingual (Italian + English) study assistant for a small group of verified classmates. It answers questions, writes summaries, and generates quizzes grounded in your course materials, citing the exact file + slide/page so answers can be fact-checked at the source.

• Grounded + cited: every claim maps to a closed-set source id rendered from trusted retrieval metadata — a citation can't point at the wrong place.
• Cross-file topic gathering via a lightweight concept index.
• Open-source models, free to run: a free Hugging Face Space + open LLMs served by Groq (→ Cerebras failover). Short retrieved snippets are sent to a no-training provider to generate answers; original files stay in a private dataset.
• Verified access only: the whole app is gated by a salted-hash login.

How it fits together

offline ingestion (this repo's CLI)            free HF Space (this app)
materials → clean text + exact anchors         pull index from private dataset
        → chunks → concept index → FAISS   ──►  retrieve + cite + Groq/Cerebras
        → upload to PRIVATE dataset             gated by login; serve downloads

Status (Phase 1)

Working now: gated login · ask questions · summaries · quizzes (all grounded, cited, Italian + English) · open/download the exact cited source next to each answer · cross-file topic gathering · tamper-proof citations (the model can only cite retrieved snippets; the file + slide/page label is rendered from trusted metadata, so it can't point at the wrong place).

Deferred to Phase 2/3 (not built yet): answers from web/online sources; OCR for scanned/handwritten materials (digital docs work today — image-only pages are detected and skipped); in-app uploads by classmates (the owner adds materials via the dataset + GitHub Action); a richer concept graph; page-level deep-links in citations (today you open the cited file and go to the stated slide/page yourself). Materials are normalized to clean text (not literal .md files).

See docs/DEPLOY.md for the step-by-step deployment guide, and docs/superpowers/specs/2026-06-09-studyhub-design.md for the full design.

Materials and secrets are never in this repo: originals live in a private Hugging Face dataset; tokens/keys/users live in Space Secrets.