ResourceRec: A CPU-Optimised Semantic Research Paper Recommendation System Using Sentence-Transformer Embeddings and Hybrid Relevance Scoring

The rapid growth of scientific literature has made it increasingly difficult for researchers to discover papers relevant to a specific query using keyword-based tools, which struggle to capture semantic relationships between differently worded but conceptually related work. Transformer-based semantic search addresses this gap, but existing systems are almost universally developed and benchmarked assuming GPU-accelerated infrastructure, placing them out of reach for researchers on standard consumer hardware. This paper presents ResourceRec, a semantic research-paper recommendation system built specifically for CPU-only inference on commodity hardware (Intel Core i5-6200U, 12 GB RAM, no CUDA-capable GPU). ResourceRec encodes paper abstracts using the all-MiniLM-L6-v2 Sentence Transformer and ranks candidates using a hybrid score combining semantic similarity, publication recency, and citation impact. On a 5,000-paper, 40-topic multi-source dataset evaluated against TF-IDF, BM25, and the larger all-mpnet-base-v2 model, ResourceRec achieves an NDCG@5 of 0.744 and a Precision@5 of 0.588 — a 9.5% relative improvement in Precision@5 over TF-IDF — while running queries 6.4 times faster and constructing its index 7.1 times faster than all-mpnet-base-v2, with no GPU acceleration required. A new ablation study isolating the hybrid re-ranking layer’s recency and citation signals found that neither differs significantly from semantic-only retrieval on its own, but their combination in the full hybrid model significantly underperforms semantic-only retrieval (paired Wilcoxon test, p < 0.05 on both NDCG@5 and P@5), an outcome traced to sparse citation and publication-year metadata rather than to the scoring formula itself and reported here as an honest negative result. These results show that transformer-based semantic literature search does not require expensive infrastructure to be practical, and that a modest, deliberate quality trade-off can buy an order-of-magnitude improvement in accessibility.