Abstract

Information Retrieval (IR) development shows a trajectory from strict symbolic systems Boolean Retrieval Model (BRM), Term Frequency-Inverse Document Frequency (TF-IDF), and Vector Space Model (VSM) towards flexible hybrid approaches using statistical BM25, Latent Semantic Analysis (LSA), and PageRank (PR)) alongside neural advancements (Neural IR (NIR) and Learning-to-Rank (LTR)) and Reinforcement Learning (RL) techniques. Traditional models focus on efficiency but fail to provide semantic flexibility, which NIR overcomes by demonstrating state-of-the-art effectiveness (NDCG@10 ≈ 0.89), although it requires significant computational resources. Hybrid systems bridge these gaps: The BM25-transformer-LTR ensemble provides high accuracy and scalability (NDCG@10 ≥ 0.85) while federated learning achieves a 30% reduction in computational overhead. Graph-based methods boost domain interpretability in fields with complex entities through spectral clustering and temporal analysis, while reinforcement learning provides personalized experiences at the cost of potential bias amplification. Ongoing major challenges span high resource demands for transformers, scaling issues with graph models, and ethical issues in designing rewards for reinforcement learning. Semantic precision will be enhanced through quantum-inspired indexing, decentralized architectures will protect privacy, and fairness-aware ranking systems will reduce bias in future developments. This research unifies heuristic rigour with neural semantic analysis and ethical responsibility to establish IR as a field where technical advancements support social obligations and deliver scalable, equitable tools that benefit healthcare and legal research.