LinkedIn has introduced an AI-powered people search feature that allows users to make natural language queries, enabling more semantic and intent-driven search results. Unlike traditional keyword-based searches, this system understands related concepts, improving the relevance and usefulness of connections surfaced. The company developed a replicable ‘cookbook’ approach involving a multi-stage pipeline of distillation, co-design, and optimization to successfully scale this technology for over 1.3 billion members.
The process began with a small, meticulously scored golden data set used to generate synthetic training data. This supported the training of a 7-billion-parameter policy model that guides relevance but is too slow for production. Through model distillation, a smaller, more efficient 220-million-parameter model was created to balance speed and accuracy. To manage the massive scale, LinkedIn shifted search indexing from CPUs to GPUs and implemented an LLM-powered summarization model that reduces input size by 20x, resulting in a 10x increase in ranking throughput.
LinkedIn emphasizes pragmatism over hype, focusing on optimizing recommender systems rather than chasing agentic AI. The AI people search tool acts as a smart tool rather than an agent, allowing future agent applications to leverage it effectively. The lessons highlight the importance of focusing on one vertical at a time, codifying successful processes, and relentless optimization for enterprise AI deployment at scale.
