adidas turns customer reviews into insight at scale with GenAI

Overview

adidas built a RAG chatbot on the Databricks Data + AI Platform using AI Search, Model Serving, Unity Catalog and MLflow to analyze over 2 million customer product reviews with LLMs including Claude Haiku, cutting response latency 60%, reducing compute costs over 90%, and improving analyst efficiency 30-40%.

The challenge

adidas' existing infrastructure couldn't support GenAI-driven review analysis: their legacy chatbot wasn't built on GenAI or RAG, leading to generic answers, high compute costs and frustrating 15-second response times, while review analysis remained largely manual and nontechnical users struggled to access insights, with query payloads exceeding 200,000 tokens overloading back-end systems.

The solution

adidas embedded over 2 million product reviews using models like Databricks BGE Large, indexed them with Databricks AI Search, and deployed a RAG pipeline using Model Serving that retrieves relevant review snippets and generates responses with LLMs such as Claude Haiku; Unity Catalog secures access and governance of models and data, and MLflow tracks model performance and iteration.

Reported business value

adidas cut latency by 60% (from 15.5 to 6 seconds), reduced token input size by 98.5% (from 200,000 to 3,000 tokens per query), lowered compute costs by over 90%, and improved review analysis efficiency by up to 30-40%.