LinkedIn has re-engineered its job search platform to move beyond keyword matching, creating a system that understands complex, natural language queries. This AI-powered engine allows users to describe their ideal job in conversational terms, which the system interprets to deliver highly personalized and semantically relevant results. The architecture is built on a two-step process: retrieval and ranking. A foundational "teacher" LLM is trained to accurately rank query-job pairs, and its knowledge is distilled into lighter, more efficient models for real-time use, balancing accuracy with performance.

To ensure the system aligns with business goals, it uses multi-objective optimization, simultaneously solving for textual relevance, predicted user engagement, and the value of the match, such as the likelihood of being hired. Recognizing that real-world click data is insufficient, LinkedIn supplements its training data with millions of synthetic examples generated by LLMs and graded by human evaluators, enabling rapid model improvement.

The query engine uses Retrieval-Augmented Generation (RAG) and a "Tool Calling" pattern to enrich user queries with profile data and other context, providing personalized suggestions to refine the search. This entire system is supported by a robust GPU infrastructure designed for scalable inference, processing millions of queries per second.