3骆昱宇.DeepEye-站式数据智能体系统架构解读.pdf

1、DeepEyeDeepEye:A Steerable Self-driving:A Steerable Self-driving Data Agent SystemData Agent System骆昱宇骆昱宇香港科技大学（广州）香港科技大学（广州）https:/luoyuyu.vip/https:/ Introductionp Hierarchical Taxonomy for Data Agents(L1-L5)p DeepEye:A Steerable Self-driving Data Agent System(L4)p L4-L5:Vision of Proactive and Ge

2、nerative Data Agentsp ConclusionOutline3Tackling data-related tasks can be demandingIntroductionThe Dawn of Data AgentsData ManagementData PreparationData AnalysisTime-consumingLabor-intensiveKnowledge-heavyLLMs and LLM agents are bringing us closer to this visionLong-standing aspiration in data sci

3、ence and analytics:developing an intelligent agent capable of autonomously managing,preparing,and analyzing data to deliver trustworthy insights with minimal human intervention.4IntroductionThe Dawn of Data Agents1 ZHU Yizhang,et al.A Survey of Data Agents:Emerging Paradigm or Overstated Hype?https:

4、/arxiv.org/pdf/2510.23587Data AgentData Agent:a comprehensive architecture designed to orchestrate the Data+AI ecosystem,which autonomously addresses a wide spectrum of data-related task 1Data-relatedTasksData ManagementoConfig Tun.Query Opt.Sys.Diag.Data PreparationData Clean.Data Integ.Data Disc.D

5、ata AnalysisStruct.TableQA,NL2SQL,NL2VISUnstruct.Report Gen.5IntroductionData Agents vs.General LLM Agents AspectGeneral LLM AgentsData AgentsPrimary FocusTask and Content Centric:Completing defined tasks or generating content.Data-Lifecycle Centric:Data management,preparation,and analysis.Problem S

6、copeSelf-contained and Static:Acts on explicit instructions and a finite prompt.Exploratory and Dynamic:Actively explores and navigates vast,dynamic data lakes.Input DataSmall-Scale and Ready-to-Use:Typically receives manageable,clean inputs.Large-Scale and“Raw”:Designed to handle heterogeneous,dyna