Secure Q&A over Sensitive Data (Privacy-preserving LLM→SQL)

Stack: Python • Flask • SQLAlchemy • Postgres (or SQLite) • LangChain • OpenAI (or other LLM)
Features: schema-aware LLM→SQL translation (with conservative fallback), RBAC & PII masking, audit logging, per-user privacy budgets, Laplace & Gaussian differential privacy for aggregates.

Quick start (local, SQLite)

1. copy .env.example → .env and edit if needed
2. (Optional) create a Postgres DB and set DATABASE_URL in .env
3. install:

   python -m venv venv
   source venv/bin/activate
   pip install -r requirements.txt

4. init DB & seed:

   python3 init_db.py

5. run:

   python3 app.py

How it works

• LLM→SQL: uses LangChain's SQLDatabaseChain if OPENAI_API_KEY is set; otherwise falls back to a conservative rule-based translator that only supports basic patterns.

• RBAC & PII masking: policies determine which columns each role can access and whether PII values (ssn/name) are masked.

• Differential privacy: aggregates use Laplace (ε-DP) or Gaussian (ε,δ-DP) mechanisms. Gaussian sigma computed with sigma >= sqrt(2 ln(1.25/delta)) * sensitivity / epsilon.

• Privacy budgets: each user has an epsilon budget tracked in DB. Aggregate queries deduct the requested epsilon.