AI / LLM / RAG · 2025
Production RAG chatbot for IAT Networks — 0% injection bypass.
A fully productionised RAG chatbot deployed for IAT Networks, a regional ISP. The retrieval pipeline chains query expansion through MiniLM embeddings into ChromaDB (top-8 retrieval), followed by cross-encoder reranking to surface the top-4 most relevant chunks before handing off to Groq's SSE streaming API. A 4-layer GuardRail stack — injection regex, PII detection, domain filter, and toxicity gate — ensures no adversarial prompt escapes. Dockerised on Railway with zero downtime across all deployments.
Query expansion → MiniLM embedding → ChromaDB top-8 retrieval pipeline
Cross-encoder reranking to top-4 for precision context delivery
4-layer GuardRail: injection regex, PII filter, domain filter, toxicity gate
Groq SSE streaming responses — <800ms time-to-first-token
Dockerised on Railway with zero-downtime continuous deployment
0% injection bypass rate across full adversarial test suite