Retail support RAG service

Ingests product catalogs, SOPs and past tickets into a vector store, serving a grounded Claude assistant that answers retail ops queries with sources attached.

Engineering · Stack: Python, Claude, Pinecone, FastAPI, Git. Outcome: ~55% fewer escalations to senior engineers

Overview

A grounded support assistant that answers retail operations questions from the company's own catalogs, SOPs and ticket history — with sources attached.

The problem

Frontline retail staff escalate routine questions to senior engineers because the answers are buried across product catalogs, SOPs and old support tickets.

How it works

• Ingests product catalogs, SOPs and resolved tickets, then chunks and embeds them into a Pinecone vector store.
• A FastAPI service retrieves the most relevant context for each incoming query.
• Claude answers grounded in that context and cites the source document.
• Low-confidence or out-of-scope queries are routed to a human with the gathered context attached.

What it does

• Vector store over catalogs, SOPs and ticket history
• A source citation on every answer
• Confidence-gated escalation to a human
• FastAPI endpoint that drops into existing support tools

Architecture

A Python/FastAPI service with Pinecone for retrieval and the Claude API for grounded generation, versioned in Git.

Results

• ~55% fewer escalations to senior engineers
• Consistent, sourced answers for frontline staff
• Senior engineering time freed for real work