Transparent policy analysis

PoliViz: source-grounded AI for difficult policy documents.

A research-oriented privacy-policy assistant that retrieves relevant clauses, answers questions in plain language and keeps source evidence, confidence and risk visible to the user.

PythonFastAPIFAISSLLMsBrowser extensionElectron
PoliViz illustration showing policy text analysed with a risk score and source-grounded finding

The problem

Privacy policies are often lengthy, legally complex and written for compliance rather than comprehension. Users may accept terms without understanding data collection, sharing, retention or rights.

System concept

PoliViz detects or accepts a policy document, segments the text and uses embedding-based retrieval to identify clauses relevant to a user’s question. A language model produces a plain-language response, but the interface keeps supporting snippets and confidence visible.

Detect

Identify a policy on the current page or load a document for analysis.

Retrieve

Use sentence-transformer embeddings and FAISS to find relevant source passages.

Explain

Generate a concise answer with highlighted clauses and accessible language.

Review

Present confidence, risk indicators and source evidence for human judgment.

Transparency by design

The project avoids presenting the model response as an unquestionable conclusion. Source snippets, provenance and uncertainty are part of the primary interface, supporting verification and critical review.

Potential applications

  • Helping consumers inspect data practices before accepting terms.
  • Supporting comparative policy research across services.
  • Assisting public-policy teams with structured clause discovery.
  • Exploring visual argument maps for claims, evidence and interpretation.

Engineering direction

The backend uses Python and FastAPI, with FAISS for vector search. The interaction model is designed to work in both a browser extension and an Electron desktop environment, enabling analysis close to the user’s browsing context.