Customer Support Agent v2

Goal

Replace v1 (rule-based zendesk macros) with a retrieval-grounded LLM that handles tier-1 inbound and escalates cleanly via the Incident Response Handoff procedure.

Status

• Index public knowledge base via this vault → chunks.jsonl (see Use chunks.jsonl as the canonical RAG substrate)
• Wire up Claude with agent-memory mcp --site ./site for live read access
• Eval suite: 200 historical tickets, target 70% auto-resolve
• Add tone-of-voice constraint per [redacted-link]
• Ship behind feature flag support_agent_v2

Eval results (last run)

Risks

• Public knowledge base is incomplete in the billing area; see Data Pipeline Refactor for the fix.
• Privacy: customers' email addresses arrive in transcripts. We use --mode redacted for all chunks we re-index from transcripts (see Separate Private and Public Memory).

Related

• Weekly Agent Memory Review
• Use chunks.jsonl as the canonical RAG substrate
• Frontend Agent Handoff — 2026-05