LinkedIn Researchers Propose Unified SLM Framework for Industrial Semantic Search Query Understanding

Summary

This paper presents a unified structured query understanding framework for industrial semantic search, developed and deployed at LinkedIn. The authors propose consolidating multiple task-specific query understanding components into a single Small Language Model (SLM) using schema-constrained generation. To address data bottlenecks, they introduce Query Illuminator, a dual-purpose framework serving as both a teacher model for auto-annotation/distillation and a surrogate judge for scalable evaluation. The approach was validated through offline and online tests within LinkedIn's Job Search system, with a cross-domain case study on People Search. Results show improved user engagement and reduced operational costs while meeting strict low-latency constraints on limited GPU resources.