Discovery Omnia: Dynamic RAG for Enhanced Patent Analysis and Systematic Innovation

Access to structured and reliable knowledge sources is a crucial step in any innovation process. While general-purpose Large Language Models (LLMs) excel in language generation, they often struggle with factual consistency and precision in specialized technical domains. This research investigates how integrating LLMs with algorithmic routing and structured databases, particularly in the field of patent analysis, can mitigate issues related to hallucinations and improve targeted knowledge retrieval. This information can then be leveraged to carry out specific strategic tasks, such as extracting Patent Intelligence data to derive business insights and support decision-making, conduct a prior art search to secure a patent FTO, enhance Technology Transfer in-domain or cross-domain, but also boosting problem-solving activities and systematic innovation with TRIZ.
This research focuses on the development of a hybrid framework where natural language queries are dynamically reformulated through a routing mechanism, optimizing knowledge retrieval. This method improves the accuracy and relevance of extracted information by systematically addressing query categorization and structured search methodologies. To be able to answer such diverse goals, it is not enough to use a conventional RAG structure, but some hints will be shown on how to design a “Dynamic RAG” nowadays, which means that it is populated each time according to a complex system of translating, analyzing, expanding and routing the user question, select the appropriate source and query to retrieve the information needed and construct the final answer by AI.