Peter Smith
⬤ LinkedIn released a major upgrade to its people search functionality with an AI-driven system that understands natural-language queries and connects them to relevant skills and fields. The new search stack can process broad queries like "curing cancer" and identify oncology and genomics experts while considering whether users can actually reach these profiles through their network. The platform's 1.3 billion users now benefit from a ranking architecture built to improve relevance and make professional discovery smoother.
⬤ LinkedIn's engineering team started by creating a small "golden" dataset of several hundred to roughly 1,000 query-profile pairs scored against detailed criteria. This dataset generated synthetic examples to train a 7-billion-parameter policy judge, which was then compressed into a 1.7-billion-parameter relevance teacher alongside separate models tracking user actions like connecting and following. The final student model was trained using KL divergence on soft scores. An 8-billion-parameter retrieval model casts a wide net before a compact student ranker fine-tunes results for precision.
⬤ The system includes efficiency improvements to make large-scale deployment practical. LinkedIn reduced the ranking student model from 440 million to 220 million parameters with less than 1 percent relevance loss, making global-scale ranking more affordable. Indexing moved from CPUs to GPUs due to the people graph's unique load pattern. A reinforcement-learning summarizer cuts input context by 20×, and combined with the smaller ranker, delivers a 10× throughput gain. An LLM router evaluates queries and decides whether to use semantic retrieval or fall back to lexical search.
Peter Smith
