How Vinted Serves Personalised Search Autocomplete

22 Apr 2026

Problem

A marketplace with many countries and languages needs autocomplete that answers on every keystroke in milliseconds. The pool of suggestions is huge (125 million at Vinted), users make typos and skip accents, and a ranking built on average market behaviour shows everyone the same list. Short prefixes of 1 to 3 characters are the hardest, because generic popularity cannot tell what an individual wants. The effect on sales is also hard to measure.

Example from the article: two UK users type "sh". The baseline ranking shows both "shoes, shirt, shein, shorts". After personalisation, a user who clicks men's items sees "shoes men" first, and a user who browses women's items sees "shein", "shoes women".

Solution in brief

Generate candidates offline from product metadata and popular search queries and score them with a weighted heuristic. At query time, match prefixes through edge-ngrams indexed in Vespa, fall back from exact to fuzzy matching in tiers, then re-rank the top 20 exact matches with a LightGBM LambdaRank model that uses user features. Judge changes by engagement metrics, using interleaving when A/B tests are too slow.

Questions this case answers
  • How to rank a very large pool of candidate completions per keystroke within a tight latency budget while adapting the order to each user
  • How to evaluate many ranking variants faster than a classic randomized test allows when the business metric is far from the changed surface
  • How to build a pool of autocomplete suggestions from product metadata and search logs?
  • How to match typed prefixes against 125 million suggestions with typo tolerance in tens of milliseconds?

Tools and methods

  • Vespa — stores suggestions, matches prefixes and runs two-phase ranking
  • LightGBM (LambdaRank) — second-phase personalised re-ranking
  • Edge-ngram indexing, fuzzy matching — fast prefix match with typo tolerance
  • Team-draft interleaving, A/B tests — online evaluation
  • Apache Flink, Apache Kafka, BigQuery — streaming index updates
Continue reading — it's free

Google or a one-time email code. No password.