What you see is what you mean; intent based ecommerce search
summary
Otto integrates an LLM upfront to turn a query into intent clusters that drive retrieval, reducing backend fan‑out from 50→5 queries and preserving context across 20M products for 12M customers. Architecture uses Solr + semantic search, a Rust blending engine and neural re‑ranker; Gemini handles intent and a fine‑tuned German BERT‑style model (trained on ~300GB of product text, 1M intent‑product pairs) achieves F1 0.779 for binary match filtering. Production swaps keyword intents for facet mapping, adds conversational intent refinement and an automated virtual customer test rig to detect regressions.
AI-written summary · about
chapters (17)
- 0:00Introduction and team
- 1:07Current search architecture
- 2:35Problem with keyword-based search
- 3:41First prototype with LLM keywords
- 5:09Clustering intents for better results
- 6:56Adjusting results with LLM filtering
- 9:03Live traffic challenges and classes
- 13:24Query class analysis and caching
- 16:17Semantic distance and intent mapping
- 18:24Short intents and fewer search queries
- 20:36Fine-tuning a model for adjustment
- 22:01Formalizing selections with facets
- 25:15Shifting adjustment to facets
- 26:18Conversational search with channel back
- 29:03Tool sets and LLM control flow
- 30:49Virtual customer for evaluation
- 32:17Q&A session
related talks
51mQuery Understanding With Large Language Models: Techniques and...
28mKeynote: The Next Generation of Enterprise Search: Intent‑Aware, Hybrid, and Open
20mBeyond Relevance: Using LLMs to Transform E-commerce Search into Product Discovery Engines
30mBeyond Keywords: Building an AI-Powered Search That Reads Your Mind
01h 30mUsing LLMs to improve your Search Engine
22mAI-Powered Personalized Discovery with Graph Search and VLM-Enhanced Embeddings
22mAgentic Hybrid Retrieval: Orchestrating LLM Agents, Hybrid Search & Explainability in OpenSearch
47mBoosting AI With OpenSearch: Vector Stores, Similarity Search, and Retrieval-Augme...