Mphasis Data Analysis Agent Design Overview

This technical report presents the design of a data analysis agent developed by Mphasis NEXT Labs, focusing on handling diverse data sources and complex analytical queries. The document outlines the challenges associated with reasoning models, particularly in the context of data analysis, where effective solutions must address heterogeneous data sources and support open-ended analytical queries. The approach involves creating a workflow of sub-agents, each responsible for specific tasks in the reasoning and planning process. Key tenets include progressive abstraction of information and multi-step refinement to enhance accuracy and relevance in planning. The report details the agent's building blocks, including goal construction, contextual reasoning, and adaptive planning. Results from evaluations on benchmark datasets, DABstep and DABench, demonstrate the agent's performance, highlighting areas of success and opportunities for improvement. The findings indicate that while the agent excels in planning and execution, there are inconsistencies in handling specific data scenarios, particularly with null values.