Accelerating Contextualization in AI Large Language Models Using Vector Databases
Raad Bin Tareaf, Mohammed AbuJarour, Tom Engelman, Philipp Liermann, Jesse Klotz · 2024
In this study, we focus on evaluating the performance of machine learning models, specifically, language models, in interpreting and understanding a subset of Berlin Parliament textual documents. With the ultimate goal of identifying the most effective and reliable model, we conduct comparative analyses involving OpenAI’s gpt-3.5-turbo and our in-house model. The evaluation framework incorporates Vicuna-13B, a refined model developed through fine-tuning and specifically designed for complex performance assessment of chatbots. Despite limitations tied to the scope of the dataset and hardware constraints, our results indicate that both models exhibit consistent performance metrics, suggesting their practical applicability in real-world settings. Moreover, the study emphasizes the potential of incorporating vector databases for contextual data retrieval as a future avenue for enhancing the efficiency and understanding of language models. The paper concludes by proposing specific optimization pathways for future work, particularly in terms of software parameter tuning and model fine-tuning, to elevate the capabilities of existing models.