I built three tightly‑coupled production engines that let the Hiver AI Support System automatically handle Apple‑support style queries while keeping safety and compliance front‑and‑center.
The intent classifier uses a hybrid semantic centroid model sentence-transformers/all-MiniLM-L6-v2.
If the top‑intent confidence is below 0.60 the query is forced into OUT_OF_SCOPE_AMBIGUOUS for safe human escalation.
The taxonomy consists of five canonical classes: OS_SOFTWARE_TROUBLESHOOTING, HARDWARE_AND_BATTERY, ACCOUNT_BILLING_ICLOUD, HOW_TO_CONFIGURATION, OUT_OF_SCOPE_AMBIGUOUS.
The RAG engine indexes genuine @AppleSupport conversation pairs in a chromadb vector store.
For each query it retrieves the top k = 3 intent‑filtered historical resolutions.
Drafted replies must be ≤ 280 characters to fit a single Twitter tweet.
Only official Apple links (apple.co/... or support.apple.com/...) are allowed; any other URL is rejected.
The deterministic triage gate runs six priority checks (thermal hazards, PII/security, human‑agent request, fraud/frustration, model uncertainty, guardrail status) before deciding auto‑handle versus escalation.