This paper presents a novel method, "MetaGen Blended RAG," to address the challenges faced by Retrieval-Augmented Generation (RAG) on domain-specific datasets: isolated behind firewalls and rich in complex and specialized terminology not encountered during LLM pretraining. To address three key challenges of existing RAGs—interdomain semantic variation, the cost of fine-tuning and lack of generalization, and the difficulty of achieving zero-shot accuracy—we propose a method to enhance semantic retrieval through a metadata generation pipeline and a hybrid query index utilizing dense and sparse vectors. By leveraging key concepts, topics, and abbreviations to generate a metadata-rich semantic index and an enhanced hybrid query, our method achieves robust and scalable performance without fine-tuning. It outperforms existing zero-shot RAG baseline models on the PubMedQA, SQuAD, and NQ datasets, and even competes with fine-tuned models. This represents a novel approach to building semantic retrieval systems with superior generalization across domains.
