Modern architectures for information retrieval and recommender systems

Abstract

This paper explains, in simple words, how modern search and recommendation systems work. We start from keyword matching (BM25) and then describe vector‑based methods (dual‑encoders), token‑level matching (ColBERT) and re‑ranking. We also show why fast vector indexes like FAISS and HNSW are needed to search large collections. For recommendations, we outline the two‑stage design (candidate generation → ranking) and neural models such as Neural Collaborative Filtering. We finish with evaluation notes and practical trade‑offs.