OPTIMIZING RETRIEVAL-AUGMENTED GENERATIONFOR AI-DRIVENHEALTHCARE PATENTS VIA HYBRID SPARSE-DENSE FUSION ANDCROSS-ENCODERRERANKING

Abstract

Patent similarity search is a critical enabler of prior-art discovery and technology trend analysis in the fast-evolving domain of Artificial Intelligence in healthcare. A tension between relevance and explainability is found in existing pipelines, however. Clas sical lexical approaches using Term Frequency–Inverse Document Frequency (TF IDF) capture terminological variation efficiently but fail under paraphrasing and struc tural change. Advanced neural methods with dense sentence-transformer embeddings achieve strong results in semantic matching, but are opaque systems that lack the struc tured rationale required in legal and intellectual-property settings. This has led to a pressing demand for computational tools that are both highly accurate and fully inter pretable. This thesis introduces a multi-stage unified framework for AI-driven healthcare patent similarity analysis, which combines a hybrid sparse-dense fusion mechanism and a cross-encoder reranking stage to optimize the Retrieval-Augmented Generation (RAG) technique. Given a corpus of 18,109 cleaned patent documents from the Har vard USPTO dataset, the system first combines TF–IDF sparse vectors with dense sentence-transformer embeddings to model both terminological and conceptual relat edness. The fused hybrid representation is indexed with FAISS for efficient first-stage retrieval. A cross-encoder model jointly encodes query-candidate pairs to produce a highly precise reranked list.These top reranked patents are then passed into a RAG module that uses Google Gemini to generate natural language explanations, binary patentability answers, and quantitative similarity scores based on the retrieved evi dence that are explicitly grounded in the retrieved evidence. Standard metrics in information retrieval (IR) are employed for the evaluation of the proposed framework such as Normalized Discounted Cumulative Gain (nDCG), Mean Average Precision (MAP) and Recall at different values of k. The experimental results show that the hybrid fusion improves the recall performance over sparse baseline and dense baseline, and the cross-encoder reranking has significant improvement on top-k precision, the full pipeline achieves an nDCG@10 of 0.86. The RAG module offers clear and effective rationales and actions for similarity judgments. As a whole, this end-to-end pipeline is a powerful, transparent and efficient prior-art analysis tool for the evolving and complex landscape of AI-based healthcare patents.