The Problem
Community platforms rely on rigid, keyword-based search bars that completely fail on complex queries (e.g., 'Android developers near Lucknow'). Connecting an LLM directly to the database to solve this risked catastrophic data leakage and SQL injection attacks.
The Hard Constraint
We needed a 100% defense against adversarial prompt injection scenarios while securely querying over 1,900 cross-platform records using an untrusted LLM extractor.
Architecture
Results & Caveats
The Outcome
Successfully engineered a 5-stage code-enforced pipeline. Treated the LLM purely as an untrusted JSON extractor locked by schemas. Achieved 100% defense against 31 distinct adversarial prompt injection scenarios.
Honest Caveats
The multi-stage pipeline introduces a slight latency overhead compared to a direct text-to-SQL approach, heavily dependent on the response times of the initial Llama-Guard security layer.