Using Vector Databases for Multimodal Embeddings and Search
summary
Shows how open-source multimodal encoders (example: ImageBind) plus an open-source vector database (example: we8) convert images audio video and text into 300–3,000‑dim vectors, enabling real‑time cross‑modal search (text→images/audio/video, image→audio, etc.). Demo runs locally on a laptop and on cloud, indexes 100k Wikipedia articles for retrieval, supports 100+ languages, and scales in production to tens of billions of items with ANN indexing (avoids brute‑force). Covers RAG pipelines, multimodal recommenders, and practical CRUD and reindexing details.
AI-written summary · about
chapters (10)
- 0:00Introduction to vector search
- 2:22How vector databases work
- 5:29Vector space and machine learning
- 10:06Demo: Wikipedia search
- 15:48Multimodal models explained
- 22:27Multimodal search applications
- 29:08Recommender systems
- 33:25Retrieval augmented generation
- 41:14Vision-language demo
- 45:02Q&A and concluding remarks
related talks
41mGetting Started With Vector Databases
45mA gentle introduction to Vector Databases
59mContextual search with vector search: exploring your options with open source tools
50mA fun and absurd introduction to Vector Databases
33mVector Databases for GenAI applications
24mIntroduction to Vector Search with Weaviate
48mA fun and absurd introduction to Vector Databases
06mWhat makes vector databases essential for RAG and AI apps?