Multi-Domain Support Triage Agent
Solo-built sequential multi-agent system with local RAG that triages support tickets across isolated knowledge domains — 96.6% citation grounding accuracy across 29 pilot tickets.
Support triage at scale suffers from cross-contamination — agents hallucinate answers by mixing context across unrelated knowledge domains. This project solves that with domain-scoped RAG retrieval across isolated knowledge bases, with a safe fallback when structured output fails.
Built a sequential multi-agent pipeline in Python using Gemini Flash and Google ADK, retrieval agent first, format agent second. Chunked and embedded 774 knowledge articles (HackerRank, Claude, Visa) into 6,454 ChromaDB vectors for fully local retrieval with no cloud dependency. Evaluated SBERT vs EmbeddingGemma on a 10-ticket sample, chose EmbeddingGemma for stronger citation alignment on ambiguous and multi-hop evidence despite slower speed. Validated all outputs with Pydantic schema enforcement. Missing or invalid outputs auto-fallback to a safe escalation response instead of guessing.
Processed 29 consecutive pilot tickets at 96.6% citation grounding accuracy. The sequential multi-agent separation improved reliability, made failures easier to debug, and kept schema output stable, a tradeoff discovered from ADK's unreliable single-request structured output support across model setups.