Integrating Knowledge into Distributed Agents Using the Petals Library

Our research focused on developing strategies to improve the performance of distributed agents that rely on language models. The main goal was to enhance their functionality by integrating specialized domain knowledge. We used various approaches, including machine learning methods, natural language processing, and fine-tuning of generative models, with an emphasis on the Retrieval-Augmented Generation (RAG) method. The aim of the study was to analyze and compare different methods of knowledge integration to identify new strategies for improving agent efficiency. Our experiments showed that the Prompt-Tuning (PT) method delivers more promising results compared to RAG. Models fine-tuned using PT demonstrated higher performance, and combining them with RAG further improved the metrics. The results confirm that using knowledge integration methods with Prompt-Tuning significantly enhances the performance of distributed agents. This opens new prospects for their application in various fields, such as finance, and contributes to improving the accuracy and quality of the tasks they solve. Furthermore, our research revealed that the combination of PT and RAG is particularly promising. This approach effectively integrates domain knowledge into models and enriches the context of generated responses with relevant information, significantly boosting the performance of distributed agents in specific knowledge areas.