Design a VPC-only HR policy assistant

Altamira Global Services (fictional) is the India GCC of a European insurer: 18,000 employees in Bengaluru, Pune and Hyderabad, plus 1,200 staff in Kraków and Manila who use the same HR team.

HR wants an assistant in Microsoft Teams that answers policy questions ("How many days of paternity leave do I get?", "Can I carry forward leave into next year?", "What is the night-shift allowance for my grade?").

Constraints from the client:

• No public internet egress. Everything runs inside their Azure subscription and virtual network, reached through private endpoints. Model endpoints must be in an Azure region they already use (Central India or West Europe).
• 340 policy documents (PDF and Word), many with country- and grade-specific tables. About 15 documents change each month; a superseded policy must stop being used the day the new one takes effect.
• Answers depend on the employee's location and grade, which are in Entra ID (Azure AD) attributes and the HR system.
• Every answer must cite the policy section. If unsure, it should hand off to an HR ticket, not guess.
• P95 latency under 5 seconds. Expected load: 6,000 questions a day, peaking at 9:30-11:00 IST.
• Budget: ₹6 lakh a month for model and search costs.
• HR will not accept anything that answers salary questions about other employees.

Design the system.