Hands-On with LLM-Powered Recommenders: Hybrid Architectures
summary
Builds a working movie recommender in Colab that actually retrieves candidates with sentence-transformers and a Faiss index, then sends the top‑K (usually 25) candidates to an OpenAI LLM for zero‑shot reranking and short natural‑language explanations. Uses a 384‑dimension embedding space (384-d embeddings), filters to ~450 popular movies, reports median ratings of 3, and covers engineering tradeoffs like scaling, latency, caching, distillation/quantization and fairness. Requires an OpenAI API key and a small billing balance.
AI-written summary · about
chapters (17)
- 0:00LLM recommender systems intro
- 5:06Why recommenders matter
- 7:17Classic recommendation methods
- 10:09Two-stage architecture
- 16:08Embeddings for scalable retrieval
- 18:17ANN search explained
- 20:46Zero-shot ranking with LLMs
- 22:12Hands-on tutorial setup
- 24:06Loading MovieLens dataset
- 27:22Data cleaning steps
- 29:31Exploratory data analysis on ratings
- 35:19Filter popular movies
- 39:56Movie embeddings with Sentence Transformers
- 43:08FAISS vector index
- 47:01Retrieval function implementation
- 51:10LLM reranking demo
- 55:52Q&A
related talks
30mLLM-Based Recommendation Systems
01h 31mHands on - Build a Recommender system
01h 21mUnlocking Personalization: A Deep Dive into Modern Recommendation Algorithms
01h 14mBuilding a Recommendation System with a Hybrid Search From Scratch
23mML-Powered Search & Recommendation 101
32m"You may also like"
46mBeyond Bag of Words A Practitioner’s Guide to Advanced NLP
33mNeural Networks for Recommender Systems