Criteo

Criteo uses k-nearest-neighbour search for targeted ads, to pick the most relevant products to display. Its R&D team publishes on the Criteo R&D Blog on Medium: https://medium.com/criteo-engineering

The team spent weeks calibrating FAISS index choices across query speed, index size, recall and build time. It released the result as the open-source library autofaiss.

In retail media, Criteo places Sponsored Products ads beside organic results in retailers’ own search. Its proprietary CLEPR model acts there as a relevance guardrail, checking that candidate ads match the query before performance optimization ranks them. Criteo frames this work around two targets: accuracy (does the product match the query?) and outcome-based relevance (do clicks and purchases confirm it?). The same split underlies its agentic commerce recommendation service for AI shopping assistants.

Tools

• autofaiss — automatic FAISS index selection under memory and latency budgets
• CLEPR — proprietary two-tower query–product relevance model

People

• Victor Paltz — author of the autofaiss launch post
• Romain Beaumont — co-author
• Maxime Vono — author of the outcome-based relevancy post
• Paul Coursaux — author of the CLEPR post

Published Articles

• Introducing Autofaiss - An Automatic K-Nearest-Neighbor Indexing Library At Scale
• Leveraging Commerce Data for Outcome-Based Relevancy in Agentic Recommendation Systems
• Introducing CLEPR, our model for semantic understanding

Related Concepts

• Approximate Nearest Neighbor Search · Vector Quantization · Dense Vector Retrieval
• Outcome-Based Relevance · Semantic Relevance · Bi-Encoder