RAG-powered merchandising assistant for a D2C apparel brand — Pro tier

The merchandising lead was personally maintaining a 2,100-row spreadsheet of cross-sell pairings. Return rate on recommended items had drifted to 22% (six points over store average) as the catalog grew. Two SaaS personalization vendors quoted $240k+/year on annual contracts; both required a 60-day integration before any signal.

The ask wasn't a recommendation engine — it was a tool the merchandiser could query ("what should I pair with the new linen shorts?") and trust the answer. RAG with citations into the actual catalog + clickstream beats both the spreadsheet and the SaaS demos because the merchandiser can see *why* the assistant suggested what it did.

Pro tier build, 11 days. Day 1-3: Pipeline from Shopify + clickstream warehouse into a hybrid index (BM25 + dense embeddings on catalog metadata, image alt-text, and 26 months of co-purchase signal). Day 4-7: Built the assistant as a single-turn RAG with strict citation — every suggested SKU comes back with the retrieval reason (co-purchase frequency, embedding similarity, copy-rule match). Day 8-9: Built an eval set of 180 merchandiser queries with the lead, scored top-5 retrieval relevance by hand, tuned reranker. Day 10-11: Slack integration and the Pro-tier dashboard (cost-per-query, retrieval accuracy, suggested-vs-accepted rate).

A/B framework ran a 15% holdout from week 3 through week 9, instrumenting AOV, revenue-per-session, and return rate on recommended items.

AOV in the treatment arm landed $9.80 above holdout over the six-week test. Return rate on recommended items fell from 22% to 14.6%. Merchandiser queries cost $0.018 each at scale — the team runs roughly 200/day, so the assistant costs ~$110/month against the $400/mo Pro retainer. The merchandising lead reclaimed roughly 30 hours/month previously spent on the spreadsheet. The retainer continues; we're scoping a Custom-tier build for a multi-tenant version they want to license to two sister brands.