Shoppers needed accurate product answers and recommendations — grounded in the live catalog, with no made-up specs.
Built a full MCP tool-agent over hybrid retrieval — PostgreSQL full-text search fused with pgvector (HNSW) — with claim/number validation that blocks anything not in the catalog. A structured intent classifier routes requests; answers stream over WebSocket, and an e2e/regression harness with deterministic replay keeps quality measurable.
An assistant that answers and recommends strictly from the live catalog — no hallucinated specs — with a ChatGPT-style streaming UX and measurable quality gates.
