Transforming Product Discovery with Visual Search at leboncoin
Millisecond-Level Search
Meeting the strict 200ms latency requirement.
80 Million Ads, Seamless Retrieval
Scaling vector search across a massive dataset.
From Zero to MVP in Six Months
Rapid deployment of a production-ready visual search.
Zilliz Cloud gave us the speed and scale we needed to power visual search at Leboncoin, meeting our sub-200ms latency target and making product discovery seamless for millions of users.
Yann Lemonnier
About Leboncoin
Leboncoin is one of France’s most visited websites and a leading re-commerce platform, enabling millions of users to buy and sell second-hand goods, find homes, or connect with job opportunities. With over 28 million monthly unique visitors, it is France’s #1 private sales site and the second most popular e-commerce platform. Additionally, Leboncoin leads in the automobile and real estate sectors and is a key player in holiday rentals and job listings. Over 500,000 professionals and 15% of French companies use the platform to sell, advertise, or recruit, making it an essential hub for both individuals and businesses.
By offering a vast selection of unique, used, or refurbished items at affordable prices, Leboncoin empowers users to make sustainable choices that benefit both their budgets and the planet. As part of Adevinta, it stands among the top re-commerce platforms, driving the circular economy, reducing waste, and supporting sustainability while providing economic value to users and businesses alike.
The Challenge: Modernizing Product Recommendations with Visual Search
Yann Lemonnier, an ML Engineer at Adevinta (formerly an ML Enabler, helping teams adopt AI technologies), joined Leboncoin in 2024 to build a visual search feature that would eventually roll out across all marketplaces under the leboncoin umbrella.
Leboncoin enables users to buy and sell second-hand products, where sellers manually upload product details, including descriptions, prices, and photos. This process generates a massive database, with approximately 80 million active ads that the engineering team must manage.
To attract a new audience and modernize the platform, leboncoin decided to introduce visual search as part of their product recommendation system. The goal was to improve user experience by making product discovery more intuitive and engaging.
Building the Visual Search System
The team started by researching visual models capable of identifying similar images within their marketplace. Once they selected a model, they realized they needed a high-performance vector database to store and retrieve the embeddings generated for similarity searches. Their research quickly pointed to Milvus as the leading vector database, but due to resource constraints, they opted for Zilliz Cloud, a managed Milvus solution.
The new visual search system introduced two key features:
- Find Similar Items– Users can click a button to discover similar products.
- Reverse Image Search – Users can upload a photo and search for matching items using the camera icon in the search bar.
The project started six months prior to Yann’s arrival, with the team working under a tight deadline to deliver an MVP in that same time frame. They had to quickly familiarize themselves with visual models, embeddings, and similarity matching. After assessing their infrastructure needs, the team determined that a vector database was essential for efficiently storing and querying embeddings to support the similarity search.
Why Leboncoin Chose Zilliz Cloud
Although Zilliz Cloud did not yet support an AWS region in Europe, the team chose Zilliz Cloud because it met their strict latency requirement of under 200ms, even when using the US-based region. Beyond Milvus compatibility, Zilliz Cloud offered several other key advantages:
- Monitoring & Scalability – Built-in tools simplified observability and scaling.
- Milvus Compatibility & Open-Source Transparency – The team could inspect the codebase for transparency and reliability.
- Ease of Deployment – Creating a cluster was quick and simple with Zilliz Cloud's intuitive UI.
Initially, the team faced challenges with data ingestion via Spark, but Zilliz Cloud’s Bulk Insert feature streamlined the process and made data ingestion much more efficient. In addition to product images, the team is also incorporating changed data (user interactions) into the ingestion process using custom software that converts events into Milvus upsert operations.
Results
By implementing visual search powered by Zilliz Cloud, Leboncoin successfully modernized its product recommender system, delivering a seamless and intuitive experience for users. Key outcomes include:
- Improved User Engagement: The introduction of visual search significantly enhanced user interaction by allowing them to easily find similar products. The reverse image search and “Find Similar Items” features led to increased product discovery and improved user retention.
- Fast and Scalable Performance: With over 80 million ads in the marketplace, the system exceeds the demanding 200ms latency requirement, achieving even lower latency. Zilliz Cloud’s scalability ensured the platform could handle large data volumes without compromising speed, even during peak traffic periods.
- Efficient Vector Search: The use of Milvus integrated with Zilliz Cloud allowed the team to efficiently manage vector embeddings. The simple deployment and Zilliz Cloud's Bulk Insert feature made data ingestion fast and smooth, enabling rapid development of the MVP.
- Future-Proof Solution: Zilliz Cloud’s scalability allows Leboncoin to continue innovating. The team plans to integrate advanced AI features like large language models (LLMs) for automated product descriptions and explore audio-based item searches.
Future Plans: Exploring Conversational Search
Leboncoin plans to further enhance its platform with cutting-edge AI features:
- LLM-Generated Descriptions– Automating product descriptions to improve listing quality.
- Conversational Search– Enabling users to search for items through text-based queries, powered by large language models (LLM), chat interfaces, and retrieval-augmented generation (RAG) for more accurate and dynamic results.
By continuing to leverage AI and scalable infrastructure, Leboncoin is refining the shopping experience for millions of users, staying competitive in the re-commerce space.
Follow Leboncoin’s news, articles, conference talks, and meetups hosted in their offices → https://mylnker.com/leboncoin-tech