Text to Context: How We Introduced Hybrid Search
summary
Claim: a hybrid system on top of OpenSearch + KNN plugin combines BM25, vector cosine and a product popularity score via a learned linear combiner and yields higher clicks and revenue in A/B tests. Implementation: catalog encoding with PySpark, embeddings pushed over Kafka into OpenSearch dynamic vector fields, live query encoding with ONNX-served models in Python (P95 ≈ 10 ms), MLflow for models, FastAPI, Kubernetes and Airflow. Model choice: E5 small multilingual for latency (E5 large and OpenAI were better offline but too slow). Three AB tests run; third shows positive lift after threshold and location-handling fixes.
AI-written summary · about
chapters (12)
- 0:00Introduction to hybrid search
- 1:48Problem with search results
- 3:13What is hybrid search
- 8:42Hybrid search architecture
- 13:12System challenges and latency
- 15:00Inference service details
- 17:27Training and evaluating models
- 23:49Offline evaluation findings
- 24:54Live results and A/B testing
- 28:27Team setup and iteration
- 30:12Future improvements and plans
- 33:51Q&A session
related talks
30mSearch on Steroids — A Story of Semantics
37mLexical, Vector & Hybrid Search with Elasticsearch
41mIntegrating Realtime User Behavioral Embeddings Into OpenSearch
26mAdvancing Multi-Modal Search Capabilities in Search Pipeline
36mSearch Complete
46mNatural Language & Search, Improve your Search w/ Vector capabilities
47mWhy, What, When and How of Running Diff Types of Sear...
44mVector Search Isn’t Enough