LinkedIn’s content feed, used by over 1.3 billion members, was originally powered by five distinct retrieval systems, each serving different types of content and optimized independently. Over the past year, LinkedIn’s engineering team replaced these separate pipelines with a unified system based on a large language model (LLM), dramatically improving both relevance and efficiency. This redesign involved revamping content retrieval, ranking strategies, and compute resource management. By processing rich professional context and treating users’ interaction histories as sequences of professional stories, the new system delivers more tailored feeds. Additionally, LinkedIn overcame challenges of handling numerical data in prompts and optimizing GPU usage by separating CPU and GPU workloads, ultimately reducing operational costs while enhancing performance at massive scale.
Back
