SupportBot — AI Customer Support

## Executive Summary **Project:** SupportBot — AI-powered customer support platform **Architecture Style:** Multi-Agent System with RAG pipeline **Primary Stack:** TypeScript / Next.js / React / Node.js / PostgreSQL / Pinecone ### User Roles - **End User** — interacts with the chat widget on customer's website - **Support Agent** — monitors conversations, takes over from AI when escalated - **Support Manager** — views team performance metrics, manages canned responses - **Admin** — configures knowledge base, agent prompts, analytics, billing ### Key Decisions 1. **4 AI agents in a multi-agent architecture** — Triage Agent routes to Technical, Billing, or General specialist agents 2. **RAG pipeline with Pinecone** — Knowledge articles are chunked, embedded, and stored in a vector DB for semantic search 3. **Human handoff protocol** — agents escalate when confidence < 0.7 or urgency >= 4; full conversation context is transferred 4. **Guardrails per agent** — each agent has specific boundaries (e.g., Billing Agent cannot process refunds, Technical Agent must cite KB sources) 5. **WebSocket-first chat** — real-time bidirectional messaging between widget and server ### AI Agent Details | Agent | Model | Orchestration | Tools | Guardrails | |---|---|---|---|---| | Triage Agent | Claude Sonnet | Multi-agent router | classify_intent, check_urgency, route_to_specialist, escalate_to_human | Max 2 routing hops, escalate on low confidence | | Technical Agent | Claude Sonnet | ReAct loop | search_knowledge_base, lookup_customer, check_service_status, create_ticket, escalate_to_human | Must cite sources, no guessing, redact PII | | Billing Agent | Claude Sonnet | ReAct loop | get_billing_history, explain_charge, change_plan, escalate_to_human | Never process refunds, confirm before plan changes | | General Agent | Claude Haiku | Single-turn | search_knowledge_base | Under 200 words, KB sources only | ### Risks & Mitigations | Risk | Likelihood | Mitigation | |---|---|---| | AI hallucination in technical answers | High | Mandatory KB citation + 2-search-then-escalate rule | | Routing loops between agents | Medium | Maximum 2 routing hops before forced human escalation | | PII leakage in conversation logs | Medium | Automatic PII redaction + data retention policies |