ContextMate: a context-aware smart agent for efficient data analysis

Pre-trained large language models (LLMs) have demonstrated extraordinary adaptability across varied tasks, notably in data analysis when supplemented with relevant contextual cues. However, supplying this context without compromising data privacy can prove complicated and time-consuming, occasionally impeding the model’s output quality. To address this, we devised a novel system adept at discerning context from the multifaceted desktop environments commonplace amongst office workers. Our approach prioritizes real-time interaction with applications, according precedence to those recently engaged with and have sustained prolonged user interaction. Observing this landscape, the system identifies the dominant data analysis tools based on user engagement and intelligently aligns concise user queries with the data’s inherent structure to determine the most appropriate tool. This meticulously sourced context, when combined with optimally chosen prefabricated prompts, empowers LLMs to generate code that mirrors user intentions. An evaluation with 18 participants, each using three popular data analysis tools in real-world office and R &D scenarios, benchmarked our approach against a conventional baseline. The results showcased an impressive 93.0% success rate of our system across seven distinct data-focused tasks. In conclusion, our method significantly improves user accessibility, satisfaction, and comprehension in data analytics.