Sarcouncil Journal of Engineering and Computer Sciences

Abstract: Modern enterprises invest heavily in professional development, yet employees often struggle to discover relevant courses within vast corporate learning catalogs that align with their specific career goals. This underutilization of learning platforms, where users may only interact with 40% of available software features (Userlane. 2025), stems from a fundamental challenge: the semantic gap between a user's expressed goal and the descriptive text of an ideal course. Traditional search methods frequently fail to bridge this gap, and even classic query expansion techniques like Pseudo-Relevance Feedback (PRF) risk query drift by over-relying on potentially flawed initial search results. This paper introduces a novel recommendation framework that leverages Retrieval-Augmented Generation (RAG) enhanced with a preliminary generative step based on Hypothetical Document Embeddings (HyDE). Our system first employs a lightweight "Semantic Router" to validate user intent, ensuring that only relevant learning-related queries proceed. For valid queries, the core innovation is a generative model that acts as a domain expert, transforming a user's high-level goal into a detailed, hypothetical summary of an ideal training course. This richer, more descriptive summary, rather than the original terse query, is then used as the basis for a dense vector search against the course catalog. We conduct experiments comparing our HyDE-based approach against three baselines: a traditional keyword-based search, a standard dense retrieval model, and a PRF-based query expansion model. The results demonstrate that our method achieves significant improvements in recommendation relevance, increasing nDCG@10 by 21% over the next best baseline. This work validates the efficacy of applying generative, corpus-independent pre-retrieval steps to highly specialized domains like corporate education, offering a scalable solution for fostering continuous and personalized employee growth.