Kleinanzeigen

Germany’s largest online classifieds marketplace, serving millions of users daily. Formerly known as eBay Kleinanzeigen. Operates one of the highest-traffic homepage personalized feeds in European e-commerce.

Search Engineering

Kleinanzeigen migrated their homepage feed from Elasticsearch to Vespa, moving user behavioral profile management, scoring, and retrieval inside the search engine itself. The key motivation was eliminating an external orchestration layer that required round-trips for every new behavioral signal.

Architecture Highlights

• Three document schemas in Vespa: ads, user_category_profile, relations
• WAND retrieval: sparse inner-product matching over field.value attribute tokens (e.g., color.red, brand.bmw) without full corpus scan
• In-engine profile updates: a Vespa document processor intercepts click events and updates user profiles in-process — no external profile store
• Token similarity propagation: a relations document defines similarity coefficients between attribute tokens, allowing partial weight sharing (e.g., color.red → color.blue)
• Signal decay: click-derived attribute weights decay over time and are L2-normalized

People

• Andre Charton — engineer, author of the Vespa migration writeup

Case Studies

• Kleinanzeigen - Vespa Migration for Homepage Feed

Articles

• From Elasticsearch to Vespa - Rebuilding the Kleinanzeigen Homepage Feed Part 1

Concepts

WAND · Personalization · Click Signals · Sparse Vector Retrieval · Vespa