Conception and development of design patterns for the use of large language models with retrieval augmented generation in applications with generative artificial intelligence
Due to digitalisation and the enormous increase in data volumes, organisations are faced with the challenge of identifying and using relevant information. A change is currently taking place in many industries. This is because large language models (LLMs) such as GPT-4 offer new opportunities, e.g. generative chatbots and virtual assistants in customer support.
Better answers through RAG
However, LLMs also have disadvantages: they sometimes generate outdated, incorrect or fictitious information, draw on untrustworthy sources or invent new ones. One solution to this problem is the use of “Retrieval Augmented Generation” (RAG). With this technique, LLMs can be enriched with additional sources of knowledge. This provides more up-to-date, accurate and relevant answers and reduces the error rate. However, many organisations currently lack the methods to integrate RAG into their operations. RAG also increases complexity and costs. At the same time, it can reduce hallucinations of LLMs, but not completely avoid them.
Design patterns for smooth deployment
In his dissertation, Bastian Stock wants to develop design patterns to create and optimise applications that use language models and are based on LLMs and RAGs. He will enrich these design patterns or solution templates with best practices and set them up in such a way that they meet the technical regulatory requirements for LLM applications. In the subsequent evaluation, he plans, among other things, to assess the effectiveness and efficiency of his proposals and provide insights for future applications.
