Vinted

Europe’s largest online second-hand fashion marketplace. Serves users across multiple European markets with personalized item recommendations on the homepage.

Search & Recommendations Architecture

3-stage recommender system:

1. Recall (Vespa + ANN): Two-tower model — user embedding (from interaction sequence) vs. listing embedding (brand, price, size, photos). ANN retrieves candidates in <100ms. Key requirement: prefiltering by user-set filters (size, category) at query time.
2. Downstream ranking stages

Why Vespa over FAISS: FAISS is a read-only index requiring periodic rebuild; no prefiltering. Vespa supports real-time updates and metadata-filtered ANN natively.

Benchmark results (1M docs, 256-dim):

• Vespa indexing: 3.8× faster than Elasticsearch
• Vespa query throughput: 8× higher RPS before CPU saturation
• Vespa P99 latency: 26ms vs. 110ms for Elasticsearch

ANN accuracy: ~60–70% recall vs. exact search, but user satisfaction did not improve with exact search (latency cost not worth it).

Search Platform Migration

Vinted ran on Elasticsearch from 2015 until migrating its billion-item, 20k-RPS search platform to Vespa (item search Nov 2023, facets April 2024) — halving the server fleet and cutting change-visibility from 300s to 5s. See Vinted - Migrating Search from Elasticsearch to Vespa.

Key Articles & Case Studies

• Vinted - Migrating Search from Elasticsearch to Vespa — the ES→Vespa search-platform migration
• Adopting Vespa for Recommendation Retrieval at Vinted — ANN recommendation retrieval on Vespa (Dainius Jocas, Aleksas Kateiva)
• Optimizing Vespa Latency with Match-Features at Vinted — match-features for in-engine feature serving + P99 9ms→3ms (Dainius Jocas)
• Dense Retrieval at Vinted — multilingual-CLIP two-tower hybrid semantic retrieval on Vespa (Laurynas Jasiukėnas, Dainius Jocas)