Mixture of Encoders: A Vector-Native Approach to Search
summary
Mixture of Encoders maps text, numbers, categories, images and temporal signals into a single Mixture of Encoders embedding so queries and metadata search together by approximate nearest neighbor instead of filters and rerankers. The system composes specialized sub-encoders into a vector-first concatenated vector, updates partial vectors, runs on Radius/Quadrant at millions of vectors via Kafka pipelines, uses LLM pairwise judging for benchmarks, and reports production wins (over $10M attributed, 77% conversion uplift and 68% AOV in a retailer case study).
AI-written summary · about
chapters (7)
- 0:00Introduction to talk
- 1:10Modern search expectations
- 2:38Problems with standard embeddings
- 6:13Mixture of encoders solution
- 8:41Benchmarking approach
- 10:26Production impact with SuperL
- 13:13Q&A session
related talks
33mAI-Powered Search in Modern E-Commerce
38mSearch: A New Era
57mUsing Vector Databases for Multimodal Embeddings and Search
19mVectorize Your Open Source Search Engine
59mContextual search with vector search: exploring your options with open source tools
56mApplications of Multimodal Learning in Media Search Engines
13mThe State of AI Powered Search and Retrieval
15mMastering RAG and Vector Search with Elastic