Semantic ID Based Tagging and Retrieval for Improved Job Matching

Abstract

The job search system requires the assignment of user queries to structured tags such as occupations, skills, and industries. Although large language models assist in understanding user intent, open generation may violate established patterns or deviate from the initial query. We propose a unified Semantic ID space that amalgamates search queries with ESCO terminology inside a distinct lexicon. This space is constructed via a domain-oriented encoder and adopted by all modules: the labeler selects only from valid candidates; the rewriter follows simple constraints to preserve core meaning while increasing hit rate; and the ranker uses simple signals from the same code. We evaluate our approach using a public corpus comprising ESCO data, job advertisements and anonymized CVs. The findings include increased annotation precision, reduced unusual outputs, enhanced memory with consistent purpose alignment, and ongoing enhancements in ranking quality and coverage. These results suggest that shared discrete spaces serve as a concise, reliable control layer for large models in the field of job search, thus forming the basis for future cross-task expansion.