AI-Powered Search in Modern E-Commerce
summary
Claims a production-ready pipeline that combines precompute product embeddings with query-time encoder (≈50 ms) and a multi-stage flow that sends top-N vectors to a slower reranker adding 50–100 ms reranker latency. Describes fine-tuning bi-encoders with contrastive triplet loss on historical click sessions, using LLMs/VLMs to synthesize multimodal training data for cold-start, and swapping XGBoost or cross-encoders for final ranking to optimize clicks or conversions while meeting latency targets.
AI-written summary · about
chapters (19)
- 0:00AI-powered search intro
- 2:32Traditional keyword limits
- 4:46Semantic search with vectors
- 6:35Better embeddings and data
- 8:23Query types impact
- 9:51Query understanding with entities
- 11:43Reranking for relevance
- 12:52Multi-stage retrieval
- 14:15Fine-tuning embeddings
- 16:04Optimizing rerankers
- 17:27Modern search foundation
- 18:10Conversational search rise
- 19:59Recap and future
- 21:08Q&A: Encoding and reranking
- 23:20Q&A: Model speed and deployment
- 24:01Q&A: Same queries handling
- 25:04Q&A: Personalization and data
- 26:49Q&A: Semantic vs structured search
- 30:07Q&A: ChatGPT and SEO integration
related talks
39mAI-Powered Search in Modern E-Commerce
22mMachine Learning: The Engine Behind Modern E-commerce Search
13mThe State of AI Powered Search and Retrieval
26mAdvancing Multi-Modal Search Capabilities in Search Pipeline
47mLearning to hybrid search
31mAdvanced search for your e-commerce website
23mML-Powered Search & Recommendation 101
33mNeural Search - Let's talk about quality