Carousell

Southeast Asian C2C marketplace operating in Singapore, Malaysia, Philippines, Indonesia, and other markets. Search challenge: multilingual user base with mixed-language queries, and sponsored listing relevance (Spotlight product).

Search Context

Spotlight is Carousell’s paid placement product — sellers pay to appear prominently in search. The core challenge: match sponsored listings to user queries semantically, not just lexically. “Selling my old iPhone 12” should appear for query “used apple phone.”

Key Engineering Work

Dense Vector Migration

• Migrated from pure BM25 to hybrid search with dense vectors
• Model: paraphrase-multilingual-MiniLM-L12-v2 — essential for SE Asian mixed-language queries (English + Filipino, Bahasa, Mandarin)
• 384-dim embeddings chosen over 768-dim for storage/speed trade-off at scale
• Elasticsearch dense_vector field + cosine similarity

Results

• +23% CTR on Spotlight listings (semantically relevant ads shown for semantic queries)
• -18% zero-result queries — listings found where keyword matching failed
• Multilingual model was essential: users mix English with local languages in the same query

Key Lesson

Hybrid over pure dense: BM25 still wins for exact model/SKU searches; dense wins for categorical/semantic queries.

Articles

• Migrating to Elasticsearch with Dense Vector for Carousell Spotlight — migration strategy, results

Related Concepts

• Dense Vector Retrieval · Hybrid Search · Semantic Search · Bi-Encoder