Policy-Grounded Dynamic Facet Suggestions for Job Search

Researchers propose a policy-grounded, retrieval-augmented ranking framework for dynamic facet suggestion (DFS) to improve job search on LinkedIn. Over 80% of job queries contain three or fewer keywords, making intent inference difficult. The system uses offline taxonomy curation, embedding-based retrieval, and distilled small language model (SLM) scoring to generate personalized semantic attributes in real time. Optimized via pointwise single-token scoring with batching and prefix caching, offline evaluation shows high precision, and online A/B tests indicate significant improvements.