From Text to Context: How We Introduced a Modern Hybrid Search — Ansgar Gruene, Dharin Shah

GetYourGuide implemented a hybrid search system combining traditional keyword-based search (BM25) with vector/semantic search to improve search result quality

The hybrid approach uses a multilingual E5 small model for vector embeddings, allowing search in 30+ languages while maintaining good performance

Implementation achieved P95 latency of ~200ms using CPU-only infrastructure and ONNX runtime for model serving

Architecture uses OpenSearch with KNN plugin for vector search, Kafka for data transport, and MLflow for model management

Search ranking combines three main components: vector similarity scores, BM25 keyword matching scores, and product performance scores

Model evaluation used NDCG metrics and real user behavior data to validate improvements

Team implemented dynamic templates in OpenSearch to flexibly handle vector fields and enable easy model experimentation

A/B testing showed increased user engagement and revenue after deploying hybrid search

Key challenges included handling location-based queries effectively and ensuring good results for queries without locations

Future improvements include adding category search, query understanding, and optimizing vector storage through quantization