Lightweight Small Language Models with Retrieval-Augmented Knowledge for Secure Edge-Level Patent Drafting Assistance

Abstract

The increasing complexity of intellectual property documentation has created a growing demand for intelligent tools capable of assisting inventors and researchers during patent drafting. Traditional patent preparation requires extensive analysis of prior-art documents, technical literature, and existing patent claims, making the process time-consuming and knowledge-intensive. Although large language models have demonstrated strong capabilities in technical text generation, their dependence on cloud-based infrastructures raises concerns regarding computational cost, latency, and the potential exposure of confidential invention data. This study proposes a secure edge-level patent drafting assistance framework based on lightweight small language models integrated with retrieval-augmented knowledge systems. The proposed architecture combines semantic embedding techniques, dense passage retrieval, and vector similarity indexing to retrieve relevant patent documents and technical references from localized knowledge repositories. Retrieved contextual information is incorporated into the generation process of the small language model to improve drafting coherence, factual consistency, and prior-art awareness. By deploying the system on edge computing environments, sensitive invention data remains within local infrastructures, reducing the risk of intellectual property leakage associated with cloud-based processing. Experimental evaluation demonstrates that retrieval-augmented small language models significantly enhance drafting accuracy and contextual relevance while maintaining low computational requirements suitable for resource-constrained edge devices. The results indicate that integrating retrieval-augmented knowledge systems with lightweight language models provides an effective and secure approach for automated patent drafting assistance in research laboratories, corporate innovation centers, and technology startups.