Awesome Search KG

Migrating to Elasticsearch with Dense Vector for Carousell Spotlight Search

3 min read

Migrating to Elasticsearch with Dense Vector for Carousell Spotlight Search

Source: https://medium.com/carousell-insider/migrating-to-elasticsearch-with-dense-vector-for-carousell-spotlight-search-engine-e328b16155fc
Author: Carousell Engineering

Summary

A case study from Carousell (Southeast Asian C2C marketplace) describing their migration from a pure keyword-based search engine to Elasticsearch with dense vector support for their “Spotlight” ad ranking product.

Business Context

Carousell Spotlight is a paid placement product: sellers pay to have their listings appear prominently in search. The challenge:

  • Match sponsored listings to user queries semantically, not just lexically
  • “selling my old iphone 12” listing should appear for query “used apple phone”
  • Pure BM25 fails this semantic matching

Architecture Before Migration

User query → BM25 (Elasticsearch) → keyword match → ranked results

Problems:

  • Zero results for semantic queries with vocabulary mismatch
  • Can’t match synonyms or related terms without manual synonym dictionaries
  • Sponsored listings not shown when they should be

Migration Strategy

Phase 1: Add Dense Vectors

Encode all listing titles + descriptions with a multilingual sentence transformer (important: Carousell operates in multiple Asian languages).

from sentence_transformers import SentenceTransformer
model = SentenceTransformer('sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2')
 
listing_embedding = model.encode(
    f"{listing.title} {listing.description}"
)

Phase 2: Elasticsearch kNN Indexing

PUT /spotlight_listings
{
  "mappings": {
    "properties": {
      "embedding": {
        "type": "dense_vector",
        "dims": 384,
        "index": true,
        "similarity": "cosine"
      },
      "title": {"type": "text"},
      "price": {"type": "float"},
      "category": {"type": "keyword"}
    }
  }
}

Phase 3: Hybrid Retrieval

GET /spotlight_listings/_search
{
  "query": {
    "bool": {
      "should": [
        {"match": {"title": "used apple phone"}},
        {"knn": {
          "field": "embedding", 
          "query_vector": [0.1, 0.2, ...],
          "k": 100
        }}
      ]
    }
  }
}

Results

Post-migration metrics:

  • +23% CTR on Spotlight listings (more relevant ads shown for semantic queries)
  • -18% zero-result queries — listings found where previously none existed
  • Seller satisfaction improved: more impressions for semantically relevant listings

Key Learnings

  1. Multilingual matters: Carousell users mix English with local languages (Filipino, Bahasa, Mandarin) — multilingual model was essential
  2. Hybrid over pure dense: BM25 still wins for exact model/SKU searches; dense wins for categorical/semantic
  3. Embedding freshness: new listings need embeddings computed before indexing — adds latency to seller listing creation flow
  4. Dimensionality tradeoff: 384-dim MiniLM chosen over 768-dim for storage/speed reasons at their scale

Graph View

Backlinks