Foundation Models and the Future of Information Retrieval

Information Retrieval underpins most of AI, from traditional (and semantic) search, to candidate generation in recommender systems, to the “R” in RAG. While many in the new wave of AI are just now rediscovering what’s been going on in IR for the past 50 years or so, those of us in the search space have been breaking down and recombining the key building blocks of lexical, embedding-based, graphical, and structured retrieval together and comparing with newer, purely model-based approaches (such as SPLADE, ColBERT, Mixtures-of-Logits, and encoding indexes directly into LLMs and other neural networks). In this talk, after a lightning overview of relatively recent approaches to semantic retrieval, I’ll try to give a bit of a more speculative flavor for where the industry is going, and when you should consider doing something “clever” like baking your search index into an updatable PyTorch model or using an LLM to do reranking – considering not just potential relevance or latency gains, but the operational burden comparisons between highly modular IR systems and more monolithic solutions.