A RAG assistant grounded in the live product catalog that resolves 47% of support tickets without human intervention
The same assistant produced a +18% lift in product search conversion — turning the engagement into 3× revenue uplift versus support cost savings.
Starting point— Challenge
One of Turkey's largest vertical e-commerce players (fashion + home textiles). 280K monthly support tickets, average 11-hour response time, 62% structural queries (returns/exchanges/order tracking). NPS had dropped from 32 to 19 over 18 months — leadership in panic mode.
Simultaneously, the product team had a parallel request: conversion on the /search page was declining; users searching in natural language ('coffee-colored short-sleeve polo-style tee') were getting no results.
During discovery we showed that both problems shared one solution: a single conversational layer grounded in product catalog + order data. The pilot scope was reshaped accordingly.
Approach— Approach
- step 01
Week 1-2 — Catalog + ticket discovery
Indexed 1.2M product descriptions + 2 years of ticket archive. First finding: 38% of product descriptions were a single empty line or 'no description' — this would kill RAG quality. Quick side project: bulk LLM-based description enrichment pipeline.
- step 02
Week 3 — Embedding & retrieval architecture
Hybrid retrieval: BM25 (keyword) + dense embedding (Cohere multilingual-v3) + reranker. Pure dense was insufficient due to Turkish morphology; hybrid lifted recall@10 from 71% to 93%. Reranker (Cohere rerank-3) fixed precision.
- step 03
Week 4-5 — Conversational layer
Single assistant, two modes: 'support' (orders/returns/tracking) and 'discover' (product search). The assistant switches automatically based on user intent. Function calling to order APIs, semantic search to product API.
- step 04
Week 6 — A/B pilot
5% of traffic (~140K sessions/week) routed to the assistant. Control group kept classic filter search + human support. KPIs: support resolution rate, product page conversion, average basket value.
- step 05
Week 7-8 — Production tuning + rollout
Conversion uplift reached statistical significance (p<0.01, n=420K). Assistant opened to all traffic. Return flow added: auto-generated shipping label as soon as user says 'I want to return' — link arrives in mail.
Results
"When the assistant went live, the most surprising thing wasn't the support savings — it was that cart conversion went up too. A leverage we hadn't seen before."
— Client side — CPO
Technology stack
- Cohere multilingual-v3 + rerank-3
- PostgreSQL + pgvector
- OpenSearch (BM25)
- Anthropic Claude (Haiku + Sonnet)
- Cloudflare Workers
- Next.js chat UI
- Datadog + custom evals (Promptfoo)
What came next
Personalized recommendation engine
Since the assistant has been in production, the rich signal from every session feeds the product recommendation engine. Next phase: automatic alternative suggestions when the requested size is out of stock (LLM + collaborative filtering hybrid).
Where does yours sit in this picture?
In a 30-minute discovery call we listen to your current state and share an initial read on whether a similar engagement makes sense. No commitment.