Session 9: Neural Embeddings & Semantic Search

⚠️ Please note we are currently reworking the lecture and still preparing the course materials. You will find this message wherever the course materials are being updated / not finalized. ⚠️

• Bridge: SVD/AutoRec gave us embeddings — neural nets deliver better ones (Word2Vec → Sentence Transformers)

• Inductive embeddings: any text can be embedded, even unseen items → cold start solved

• Hands-on: generate text embeddings with Sentence Transformers, build semantic search

• ChromaDB as a tool: store and query embeddings in 5 lines (Client, Collection, Add, Query)

• Contrastive training (outlook): how are embedding models trained? Similar pairs close, dissimilar pairs far

Practice: Generate text embeddings, test semantic similarity; build recommender with ChromaDB; cold-start test — invent a new movie, find similar ones