RAG Design Patterns: how Gorgias uses Zilliz for AI product recommendations | Milvus Meetup Paris

Gorgias runs AI agents for e-commerce customer support and shopping assistance. Both agents need to find the right product among millions, either the exact item a shopper is asking about or a good recommendation. ML engineers Othmane Jebbari and Mohamed Ali Fathallah trace three versions of their product retrieval system: First version: a daily cron job bulk-loaded about 8 million products from 15,000 merchants. Recommendations were up to 24 hours stale, the bulk loads swamped the vector database, and each merchant was capped at 4,000 products. Second version: updates stream in through a Kafka change-data-capture consumer as they happen, Zilliz Cloud autoscales to absorb the peaks, and a re-ranker trained on Gorgias data fixes search quality. Current version: a separate enrichment pipeline turns images into text descriptions and summarises reviews. Each product is indexed as several vectors, which they call “lookup channels”. The agent itself chooses queries and metadata filters. They finish with what’s next: retrieval as a tool the agent calls rather than a fixed workflow, evaluation tied to GMV instead of recall, and three lessons from the journey. What you’ll learn: Why daily batch indexing breaks down for live product catalogues, and how CDC streaming fixes freshness and load spikes Training a re-ranker on purchase data when vector similarity alone isn’t good enough Running enrichment (image descriptions, review summaries) as its own service that feeds both search and the LLM context Migrating a live collection to a new schema with zero downtime by running two CDC consumers in parallel Indexing several vectors per product and letting the agent pick filters Why recall gains don’t always move click or buy rates, and how A/B testing every recommender keeps you honest Chapters 0:00 Introduction 0:51 Agenda 1:26 Gorgias’s AI agents: support and shopping assistant 2:08 Two retrieval jobs: identify and recommend 2:42 The first version 4:45 Limits: staleness, load spikes and scale 6:17 The search path and the quality problem 7:10 The second version: CDC with Kafka and Zilliz Cloud 9:13 Adding a trained re-ranker 10:28 The current version: an enrichment pipeline 12:39 Zero-downtime schema migration 14:08 Lookup channels and agent-picked filters 16:36 Next challenges: retrieval as an agent tool 17:56 Evaluating against GMV 19:24 Key learnings 21:23 Q&A This talk was recorded at the Milvus Meetup Paris, hosted by Criteo. The event brought together vector search practitioners to explore the algorithms, infrastructure and production patterns behind modern AI-powered search. Find us on: Zilliz Website: https://zilliz.com Milvus Website: https://milvus.io Zilliz LinkedIn: https://www.linkedin.com/company/zilliz/ Milvus LinkedIn: https://linkedin.com/company/67143129 Milvus X: https://x.com/milvusio Milvus GitHub: https://github.com/milvus-io/milvus Try Zilliz Cloud: https://cloud.zilliz.com/signup