Terms Guard AI

title: Terms Guard AI emoji: 🛡️ colorFrom: blue colorTo: indigo sdk: docker app_port: 7860 pinned: false license: mit

Terms Guard AI

Paste any Terms & Conditions or Privacy Policy and the AI flags the clauses that are risky or unfair to you, grounded in EU GDPR rules.

How it works. Each clause is embedded with a MiniLM sentence-transformer and compared against 102 hand-labelled reference clauses (52 fair, 50 unfair) that are curated into the file gdpr_clauses.py. The article on a flagged clause is the one the hand-labelled corpus assigned to the closest reference clause.

Each citation is then anchored to the actual law, not to a search over loosely related material. The retrieval layer pulls a passage from the specific article cited, drawn from a knowledge base built from the full text of Regulation (EU) 2016/679: 99 articles and 173 recitals, fetched from gdpr-info.eu by fetch_gdpr.py and embedded by build_knowledge_base.py. The old index held 4,527 chunks of data-governance course notes, and the retrieved passage could come from a different article than the one cited. That failure mode is no longer possible: the passage the reader sees always carries the same article number as the citation.

Up to 15 flagged clauses per analysis are sent to Claude for verification through one of three routes (direct Anthropic SDK, the kie.ai gateway, OpenRouter), tried in order with retry. The response carries verification_source and verification_error, so a silent fallback is impossible. The previous build reported clauses as model-verified when no model had actually looked at them; that class of bug is fixed and a 12-case regression suite passes against the live deployment.

Files

• fetch_gdpr.py — pulls the GDPR text and recitals
• build_knowledge_base.py — chunks and embeds the law
• gdpr_clauses.py — the curated clause → article map (the citation source)
• gdpr_glosses.py — plain-English glosses, used only for matching
• test_kb_retrieval.py, test_live_retrieval.py — verification tests
• test_regression.py — the 12-case regression suite
• gdpr_kb/chunks.json, gdpr_kb/index.faiss — the law text and its FAISS index
• gdpr_kb/gloss.json, gdpr_kb/gloss.faiss — the glosses and their FAISS index
• gdpr_articles.json — the fetched GDPR text, the source for the build

Built by Taiwo Alabi for an MSc in Artificial Intelligence, National College of Ireland.

Educational project. Not legal advice.