《Cognite：2024Cognite Atlas AI? 工業智能體（Industrial Agents）權威指南（英文版）（83頁）.pdf》由會員分享，可在線閱讀，更多相關《Cognite：2024Cognite Atlas AI? 工業智能體（Industrial Agents）權威指南（英文版）（83頁）.pdf（83頁珍藏版）》請在三個皮匠報告上搜索。

1、Definitive Guideto Industrial AgentsThe Cognite Atlas AICopyright,Cognite,2024 www.cognite.ai Definitive Guideto Industrial AgentsThe Cognite Atlas AI“According to a recent ARC Advisory Group Digital Transformation,Sustainability and Technology survey,Artificial Intelligence,AI is the most impactful

2、 technology for the next five years.Cognite Atlas AIs guide to Industrial Agents is a practical place to start for digital leaders looking to make AI work in complex industrial environments.”JANICE ABELPRINCIPAL TECHNOLOGY ANALYSTARC ADVISORY GROUP“Data lakes and copilots are just the beginning.Indu

3、strial AI demands more than a one-size-fits-all approach.Cognite Atlas AI unlocks the full potential of generative AI for industry with industrial agents that can accelerate efficiencies and generate tens of millions of dollars in business impact.”PAUL GRENETCHIEF REVENUE OFFICERCOGNITEPAULA DOYLECH

4、IEF DIGITAL OFFICERAKER BP“Cognite Atlas AI enables us to use AI to enhance decision-making and improve efficiency,like with an industrial agent fine-tuned to understand unstructured technical documentation and Aker BPs equipment hierarchy.By implementing the Document Parser AI Agent,we are streamli

5、ning our equipment management process,saving thousands of hours of data-punching,and refocusing our experts on business problems that really matter to the short-and long-term success of our operation.”ContentsSection 0Introduction.06Executive Summary.08 The Four Five Things You Need to Know about Ge

6、nerative AI for Industry.10Foreword.14 The Data and AI Problem.16 Demystifying Industrial AI Agents:What We Can Learn from Iron Man.20Chapter 4AI Is the DrivingForce for IndustrialTransformation.90 4.1 Verdantix View:Industrial DataOps in 2024.94 4.2 AI Will Deliver Untapped Value for Asset-Heavy En

7、terprises.102 4.3 Democratizing Data:Why AI-Infused Industrial DataOps Matters to Each Data Stakeholder.104Chapter 5Use Cases.108 5.1 Industrial Use Cases Require a System of Engagement.110 5.2 Cognite Data Fusion:An SOE to Scale Operational Use Cases.114 5.3 Improving RCA with AI Agents and Industr

8、ial Canvas.122 5.4 Examples of Industrial AI Agents.126Section 2The Business Valueof AI.88Section 3Tools.130Chapter 6Tools forthe Digital Maverick.132 6.1 Industrial AI&Data Management Software:How to Avoid Decision-Making Pitfalls When Purchasing.134 6.2 Navigating Digital Transformation:A Framewor

9、k for Success.138 6.3 Navigating Digital Initiatives by Using Value as the North Star.142 6.4 Data and AI RFP Guide.148The Cognite Atlas AI Definitive Guide to Industrial AgentsSection 1Making Generative AIWork for Industry.24Chapter 1Industrial Agents.26 1.1 The Treacherous Path to Trustworthy Gen

10、AI for Industry.28 1.2 The Path Forward:Industrial Agents.32 1.3 The Challenges of Implementing Industrial Agents.36 1.4 The Applications and Benefits of Industrial Agents.38 1.5 Agent Orchestration and Agent Ecosystems.40 1.6 Does RAG Still Matter?.42 1.7 From RAG to CAG.46Chapter 2Large,Small,and

11、CustomLanguage Models.48 2.1 Understanding the Difference.52 2.2 LLMs and Their Application in Operations.54 2.3 So Why Do We Need SLMs?.56 2.4 Custom Language Models.58 2.5 Evaluating Large Language Models Usefulness Correctness.60 2.6 Choosing the Right Model with autoLLM.68 2.7 Performance Benchm

12、arking.70Chapter 3Semantic KnowledgeGraphs.72 3.1 Defining Knowledge Graphs.76 3.2 Knowledge Graphs and Data.78 3.3 Knowledge Graphs and AI.82Section 0 IntroductionThe Cognite Atlas AI Definitive Guide to Industrial AgentsExecutive Summary.08 The Four Five Things You Need to Know about Generative AI

13、 for Industry.10Foreword.14 The Data and AI Problem.16 Demystifying Industrial AI Agents:What We Can Learn from Iron Man.20The Four Five Things You Need to Know about Generative AI for IndustrySection 0:IntroductionExecutive Summary8Little has changed in the four key points weve been preaching about

14、 generative AI from the beginning,but there is one notable addition in this guide:industrial value is accelerated by industrial agents.Though really an extension of point four,industrial agents are a significant development that enables industrial organizations to use generative AI to carry out more

15、 complex operations with greater accuracy.We will dive into more details on industrial agents in the following chapters but,if you only read one part of this guide,let it be this:Executive Summary:The Five Things You Need to Know about Generative AI for IndustryLLMs+Knowledge Graph=Trusted,Explainab

16、le Generative AI for IndustryThis is the simple formula for applying generative artificial intelligence(AI)in industry.Combining large language models(LLMs)with a deterministic industrial knowledge graph containing your operations data makes your asset performance management intelligent and efficien

17、t.Generative AI for Industry Needs to Be Safe,Secure,and Hallucination FreeAnd with the formula above,it is.You need a complete,trustworthy digital representation of your industrial reality(i.e.,an industrial knowledge graph)for LLMs to understand your operations and provide deterministic responses

18、to even the most complex questions.The Four Five Things You Need to Know about Generative AI for Industry1.2.To Apply Generative AI in Industrial Environments,the Ability to Prompt LLMs with Your Operational Context Is EverythingThis means having a deterministic industrial knowledge graph of your op

19、erations,including real-time data.You need a solution that delivers contextualized data-as-a-service with data contextualization pipelines designed for fast,continuous knowledge graph population.3.1011While Generative AI Itself Is Undeniably Transformative,Its Business Value Is in Its Application to

20、 the Real-World Needs of Process Engineers,Field Workers,Maintenance Teams,and Other Data ConsumersInnovative AI features are only valuable in a platform that also enables simple access to complex industrial data for engineers,subject matter experts,and data scientists so they can make the right dec

21、isions at the right time.Industrial Value Is Accelerated by Industrial AgentsThese tailored,AI-powered applications are designed with an in-depth understanding of industry-and customer-specific terminology,processes,and standards.They utilize algorithms and data models specifically optimized for the

22、 patterns and anomalies typical in a particular domain.And,they can be customized to fit the unique workflows and requirements of different organizations.As such,industrial agents can offer more accurate and relevant guidance and can be scaled to accommodate the growing data and complexity of operat

23、ions as an organization expands.Industrial agents bring the power of AI and machine learning directly to the challenges and tasks unique to the industry and each unique organization.This makes them crucial in improving decision-making processes to help organizations achieve higher productivity,safet

24、y,and overall operational efficiency.5.Executive Summary:The Five Things You Need to Know about Generative AI for Industry4.Contextualize Events to AssetsE.g.connect shop orders,work orders,and alarms,to correct site,prod.line,and equipmentContextualized DocumentationE.g.connect tags in P&IDs to ass

25、ociated assets,time series,work orders,and documents,etc.Contextualized 3D ModelsE.g.filter for and visualize where work orders are located on plant1213The Data and AI ProblemSection 0:IntroductionForeword14For every one person who can speak code,there are hundreds of others who do not,especially in

26、 the industrial environments where there are numerous data types and source system complexity.Subject matter experts,field engineers,and data scientists deserve simple access to all industrial data.This requires a unique way to leverage and apply contextualized data(i.e.data that is enriched with re

27、levant information and relationships,making it more meaningful and useful for analysis and decision-making).The Data and AI ProblemTime Spenton Data ProductsWithout simpleaccess to complex industrial dataWith simpleaccess to complex industrial dataGathering dataBuilding solutionsForeword:The Data an

28、d AI Problem1716Human ContextualizationFrom ThisManual,cumbersome industrial data workflowNo simple access to complex industrial data and insights.to ThisSingle workspace for data&analytics powered by AIGenerative AI is context dependent.While generative AI has tremendous potential to make data easi

29、er to explore,understand,and use,answers are often wrong without contextualized data.However,traditional efforts to connect data from systems are manual and time-consuming,and are not capable of managing structured data at scale,much less incorporating the growing unstructured data.An efficient way

30、must be found to provide generative AI solutions with more context to enable them to provide the right answers in industrial environments.Only then can it be used to optimize production,improve our asset performance,and enable AI-powered business decisions.This is where industrial agents come in.For

31、m hypothesisManual analysisFeedbackLook up P&ID(documents)Locate technical documentation(documents)Gain actionable insightsLocate sensor data(historian)Locate equipment history(maintenance system)CreateSummarize documentsand write complex queries using natural language.ExploreAccess and add any type

32、 of data in one compostable environment+Copilot powered research.CollaborateAnnotate,tag users,share insights with seamless integration into no-code analytics.Foreword:The Data and AI Problem1819Demystifying Industrial AI Agents:What We Can Learn from Iron ManWith all the hype around generative AI f

33、or industry,it seems as if there is a new buzzword almost every day.The latest?“Industrial Agents.”But is it really a buzzword?Or is it rather a more accurate depiction of the actual endgame of data&AI for industry?Few standard definitions of this term exist yet for industry but,put simply,industria

34、l agents perform specific tasks in a human-like manner when trained with the right data and when using the right AI model and capabilities.The operational co-pilots that everyone is talking about or the chatbot you use when trying to rebook your flight are all types of AI agents.They aim to automate

35、 or simplify a specific,constrained workflow to improve the users productivity.However,the agents of today,which use limited pre-programmed logic,are no match for the Gen AI-based agents of the future.If we take inspiration from the movies,were getting closer to Iron Mans“Jarvis”assistant-a supercha

36、rged intelligent virtual agent that communicates via voice commands and helps Iron Man do his best work.While were a far cry from this type of omniscient,cross-functional intelligence,the technical building blocks and terminology exist today to start developing specific industrial agents for particu

37、lar operator domains.For decades,industrial operators have been trying to use data and AI to optimize production,minimize outage risk,streamline production,and make smarter daily decisions.Theyve used physical and machine learning models to classify types of asset failure,natural language processing

38、 to search for information,and now LLMs to analyze and summarize data and make recommendations.However,with the exception of robotic process automation(RPA)for back-office functions,the impact on factory-floor operations has been underwhelming.Why is this?First,as with any new paradigm,its taken a b

39、it of time to learn and understand whats required from a technological and process perspective.For example,in the early phases of generative AI,circa 2023,the idea was to go straight to a“Jarvis-based future”with few general agents,with broad objectives trained on large sets of data,structured and u

40、nstructured.But this made it difficult to trust and repeat the results due to the inherent hallucinations and other limitations of generative AI.Fast forward to today the more realistic scenario is the orchestration of many specific virtual agents,trained on smaller,secure,relevant data sets,designe

41、d with intuitive UX to improve workflows definitively.Despite the early learnings,whats become abundantly clear is that for industrial agents to work and be trusted in industrial domains,they need three things:1.A domain-specific task or objective2.Secure,contextualized data for this objective3.The

42、most appropriate LLM for the task at handForeword:The Data and AI Problem2021Until recently,industrial organizations did not prioritize the need for secure,contextualized data foundations,which are critical for training myriad LLMs on relevant data.Today,with support from boards of directors and exe

43、cutives,even legacy industries are investing in teams and technology to bring order and context to their vast amounts of siloed data.Second,the way users interact with digitally enhanced industrial processes has not been intuitive,making it challenging to actually improve the workflow.In flight,if I

44、ron Man was not able to speak conversationally with Jarvis and had to manually look up information with precise terms,his workflow(and mission outcomes)would not be as good.In the field,operator workflows are precise and well-established.Information must be trustworthy and accessible instantly,using

45、 handheld devices and simple commands,not lines of SQL code.Technology that doesnt offer dramatic workflow improvements does not get adopted.Heres where things get even more interesting.Gartner predicts:“Large Language Models(LLMs)will become the preferred interface to enterprise data.”This means th

46、at the effort required for users to access and refine information(once they trust the outcomes)becomes human.So even though an operator may not be able to ask their agents the same breadth of questions as Iron Man could to Jarvis,their interface to answers becomes more human and intuitive than ever

47、before-making it easy to adopt into a workflow.Putting these lessons into action Iron Man didnt build Jarvis overnight,but we can make several assumptions about what it took to make this high-impact agent.The good news?The journey for industry has many parallels.1.He started with simple access to co

48、mplex data.Whether you are trying to improve operational dashboards or introduce industrial agents-both start with an industrial data foundation that uses AI to contextualize information at scale.This is the key requirement for any modern industrial digital transformation program.If youre still stru

49、ggling in this area,here is a common starting point-that has put our customers on fast success tracks.2.He probably used a knowledge graph to contextualize all his data.In the industrial space,LLMs depend on data in context,returning higher accuracy outputs because agents can be trained on narrow da

50、ta sets based on their explicit objective.Learn more about why this matters-3.He mastered model and agent orchestration:Industrial transformation has a lot of moving pieces,and proper orchestration can make or break a program.As you can see,there are many considerations for realizing an agent-based

51、industrial future that is reliable,useful,and valuable.Where should you start?By reading this book,designed to help you start or accelerate your Gen AI journey and get you closer to your own“Jarvis-inspired”operations.Foreword:The Data and AI Problem23Section 1 MakingGenerative AIWork for IndustryTh

52、e Cognite Atlas AI Definitive Guide to Industrial AgentsChapter 1Industrial Agents.26 1.1 The Treacherous Path to Trustworthy Gen AI for Industry.28 1.2 The Path Forward:Industrial Agents.32 1.3 The Challenges of Implementing Industrial Agents.36 1.4 The Applications and Benefits of Industrial Agent

53、s.38 1.5 Agent Orchestration and Agent Ecosystems.40 1.6 Does RAG Still Matter?.42 1.7 From RAG to CAG.46Chapter 2Large,Small,and CustomLanguage Models.48 2.1 Understanding the Difference.52 2.2 LLMs and Their Application in Operations.54 2.3 So Why Do We Need SLMs?.56 2.4 Custom Language Models.58

54、2.5 Evaluating Large Language Models Usefulness Correctness.60 2.6 Choosing the Right Model with autoLLM.68 2.7 Performance Benchmarking.70Chapter 3Semantic KnowledgeGraphs.72 3.1 Defining Knowledge Graphs.76 3.2 Knowledge Graphs and Data.78 3.3 Knowledge Graphs and AI.82IndustrialAgentsChapter 1Sec

55、tion 1:Making Generative AI Work for Industry26The Treacherous Path to Trustworthy Gen AIfor IndustryYou cant avoid the buzz and excitement.Gartner is saying,“Large Language Models(LLMs)will become the preferred interface to enterprise data,”1 and almost every SaaS vendor has recently announced thei

56、r Gen AI Copilot.Who wouldnt love simple access to complex industrial data and analytics finally unlocking the data-powered enterprise?Beyond the hype,however,those working with LLMs for search or analytical query generation are being met with real-world challenges(not scripted,cool demos with extre

57、mely limited real-world value):Generating Working Queries from Natural Language That Produce Correct Results Using LLMs Is Non-Trivial.1.2.3.Existing LLMs Are Computationally Incredibly Expensive and Slow Compared to Database or Knowledge Graph Lookups.Building production solutions on LLMs will resu

58、lt in a very large cloud bill.For example,GPT-4o,which is a good model with reasonable cost,would cost 0.2 cents for a 100-token prompt and a 100-token answer,which is a very conservative estimate.Cost could easily end up at 1000 x(three orders of magnitude)higher with more complex prompts.In additi

59、on,LLMs are also very slow compared to low-latency UX expectations in todays real-time software world.Even very large knowledge graph lookups are many orders of magnitude more efficient and faster.In an era of focus on sustainability,driving up computation with LLMs looks unsustainable.(This is wher

60、e small language models can be.More on that in chapter 2.)LLMs Require Context,and Using Chaining to Mitigate Context Window Limitations Can Result in Compound Probabilities and,Thus,Less Accuracy.Without providing LLMs with context,they fail on practically all aside from creative tasks,which their

61、original training corpus(the public internet)supports well.One way to provide this context is using LLMs context window for in-context learning.When comparing the context windows of popular LLMs to average enterprise data volumes,it becomes clear very quickly that the math doesnt add up.Moreover,con

62、text window limitations will persist for years,if not forever.Cost and latency aside,growing the context window size does not linearly correlate to outcomes.Studies show3 that the attention mechanism in LLMs works differently for various parts of a long context window.In short,the content in the mid

63、dle receives less attention.On the other hand,multiple prompts with shorter context windows allow iteration and optimization of each component individually and even with different LLMs.Moreover,when designed to minimize dependencies among them,it is possible to minimize the effect of compound probab

64、ilities and even run them in parallel to reduce latency.Chaining multiple prompts,of course,adds to query volume,hence increasing cost once again.1.Source:Gartner.Quick Answer:Safely Using LLMs With an Active Metadata and Data Fabric Layer.14 August,20232.Source:The economic impact of the AI-powered

65、 developer lifecycle and lessons from GitHub Copilot.27 June,20233.Source:Nelson F.Liu,Kevin Lin,John Hewitt,Ashwin Paranjape,Michele Bevilacqua,Fabio Petroni,Percy Liang.Lost in the Middle:How Language Models Use Long Contexts.31 July,2023ChatGPT made an incredible initial impact,driving significan

66、t efficiency gains to developers with programming copilots.2 Yet,making LLMs serve non-developers the vast majority of the workforce,that is is not quite so straightforward.Using LLMs to translate from natural language prompts to API or database queries,or expecting readily usable analytics outputs

67、is challenging for three primary reasons:Inconsistency of prompts to completions:No deterministic reproducibility between LLM inputs and outputs.Nearly impossible to audit or explain LLM answers:Once trained,LLMs are black boxes.Coverage gap on niche domain areas:LLMs are trained on large corpora of

68、 internet data,heavily biased towards more generalist topics,not on niche domain areas that typically matter most to enterprise users.Chapter 1:Industrial Agents2829Lastly,certain types of queries spanning many facts are not feasible with LLMs alone.For example:“Which assets have shown heat exchange

69、r fouling after 2021?”In industry domain use cases,the contrast is even more pronounced as typically,LLMs will not have been trained on any proprietary industrial data needed to answer queries,and fitting enough proprietary industrial data into the context window is impossible.Another impossible tas

70、k for LLMs is when an answer requires real-time operational data.Ask any LLM if any condenser units in plant A have a temperature below 5 degrees C right now,and they cannot answer.The solution architecture to answering such complex queries is to use agents.Agents allow ChatGPT to call out the knowl

71、edge graphmultiple times during its processing of a query4.Understanding That LLM Solutions Are Best Assessed on Usefulness Rather Than Mathematical Truism.Gen AI Is Not a Silver Bullet,but a Terrific Pathfinder!You usually do more than one Google search to find the result you are looking for.The sa

72、me will apply to our future,wherein“LLMs will become the preferred interface to enterprise data.”4The right design approach is thereby not one that instantly produces the correct result but rather an interactive interface to facilitate the process of finding the right answer,placing the user in cont

73、rol,and using understandable filter inspection(as opposed to only showcasing the generated script to the non-coder user)so that users can review and adjust the suggested filters to find data of interest.As always,the data itself needs to be provided to the LLM-enhanced interface through a determinis

74、tic knowledge graph(more on this in Chapter 3),enabling users to narrow down to relevant parts of the knowledge graph.This interface significantly helps navigate the graph to the right data,even when it might not always directly“zoom”into precisely the node/nodes that initially contain the right ans

75、wer.4.Source:Gartner.Quick Answer:Safely Using LLMs With an Active Metadata and Data Fabric Layer.14 August,2023Chapter 1:Industrial AgentsPrompt engineering including as an interactive model is the Wild West of possibilities(and security risks!).But again,more focused instructions tend to work more

76、 robustly in practice than longer prompts.Were again back to chaining(see above).On security,prompt injections can leak data unless strong data access control is in place.With Cognite,all data retrieval is done using a users assigned credentials and thus,no user will be able to get access to unautho

77、rized data through prompt injections any more than they would through conventional interfaces.All existing access control mechanisms in Cognite Data Fusion apply to generative AI use as well.3031The Path Forward:Industrial AgentsAgents are designed to achieve specific goals and can perceive their en

78、vironment and make decisions autonomously.Agents include chatbots,smart home devices and applications,and the programmatic trading software used in finance.Agents are classified5 into different types based on their characteristics but generally,they all exhibit these key attributes:Industrial agents

79、 are designed to cater to the unique needs of a specific industry.They are specialized versions of general AI assistants focused on solving domain-specific problems with a deep understanding of the industrys context,terminology,and workflows.These agents leverage advanced technologies to provide exp

80、ert guidance,automate tasks,and offer highly relevant insights into the industry for which they are tailored.For example,Cognite has deployed a document parsing AI agent to automate equipment registration by reading technical documentation and unstructured input such as data sheets to find relevant

81、information to input into a structured form.Automating data extraction in this way is projected to save more than 10,000 engineering hours.So how do you go about building an industrial agent like our document parser?5.Source:https:/attri.ai/blog/a-complete-guide-top-generative-ai-agents-use-cases-fo

82、r-manufacturersChapter 1:Industrial Agents3233How to Build an Industrial AgentAIdentify the specific industrial application or problem the AI agent will address.BDefine clear objectives,performance criteria,and constraints.CGather domain-specific knowledge and expertise.AIntegrate the trained AI age

83、nt into existing industrial systems,such as manufacturing lines,robotic systems,or supply chain management platforms.BEnsure seamless communication and interoperability between the AI agent and other components through APIs,middleware,or custom interfaces.ACollect relevant data from various sources,

84、such as sensors,historical databases,and real-time streams.BClean and preprocess the data to remove noise,handle missing values,and ensure consistency.CAnnotate and label data if supervised learning is required.AConduct rigorous testing of the AI agent in controlled environments to evaluate its perf

85、ormance,robustness,and reliability.BUse both simulated scenarios and real-world conditions to validate the agents effectiveness in achieving the desired outcomes.ATrain the AI models using the prepared data.This involves feeding the data into the model,adjusting parameters,and iteratively improving

86、performance.BUse techniques such as cross-validation,hyperparameter tuning,and regularization to optimize the models accuracy and generalizability.ARegularly update the AI agent with new data and retrain it to adapt to changing conditions and improve performance.BImplement feedback mechanisms to lea

87、rn from real-world experiences and incorporate improvements over time.AChoose the appropriate AI models and algorithms based on the problem requirements.Common choices include machine learning models,deep learning networks,reinforcement learning,and hybrid approaches.BDesign the architecture of the

88、AI agent,which may include modules for perception,decision-making,learning,and communication.ADeploy the AI agent into the live industrial environment.BContinuously monitor the agents performance and behavior to ensure it operates as expected and to identify any potential issues.Problem Definitionan

89、d RequirementsAnalysis:Integrationwith Industrial Systems:Trainingand Optimization:Maintenanceand Continuous Improvement:Data Collectionand Preprocessing:Testingand Validation:Model Selectionand Design:Deploymentand Monitoring:15263748Chapter 1:Industrial Agents3435The Challenges of Implementing Ind

90、ustrial AgentsDeveloping and implementing industrial agents comes with several challenges that can affect the effectiveness and adoption of the industrial agent.For example,industrial agents leverage natural language to understand and write code based on published API documentation and examples.This

91、 is impossible with data lakes or data warehouses where,without a contextualized industrial knowledge graph,there are no API libraries that can be used as a reliable mechanism to access rich industrial data.A more exhaustive list of challenges involved in implementing industrial agents includes:Data

92、 Quality and AvailabilityIntegrationwith Legacy SystemsScalabilityand FlexibilityReal-Time Processing and Decision-MakingInterpretabilityand TrustChallenge:High-quality,relevant,and sufficient data is crucial for training AI models.In many industrial settings,data may be sparse,noisy,or incomplete.S

93、olution:Adopt an Industrial DataOps approach that includes data validation by design and implement a robust data contextualization engine to bring together and connect industrial data from all sources(like time series,P&ID Challenge:Many industrial environments rely on legacy systems that may not ea

94、sily interface with modern AI technologies.Solution:Invest in an Industrial DataOps platform with open APIs to bridge the gap between new AI systems and existing infrastructure.Incremental integration strategies can help in gradual adaptation.Challenge:AI models need to scale efficiently to handle l

95、arge volumes of data and diverse industrial tasks.They also need to be flexible to adapt to changing requirements.Solution:Design scalable architectures and use cloud-based solutions for computational flexibility.Modular designs can enhance adaptability.Challenge:Many industrial applications require

96、 real-time data processing and decision-making,which can be computationally intensive.Solution:Optimize AI models for speed and efficiency with performance benchmarking and autoLLM.Challenge:Industrial stakeholders may be hesitant to adopt AI solutions they do not fully understand or trust.Solution:

97、Develop interpretable AI models and provide clear explanations for their decisions.Build user-friendly interfaces and visualization tools that will help gain trust.Securityand PrivacyChallenge:Protecting sensitive industrial data from cyber threats and ensuring privacy can be challenging.Solution:En

98、sure no proprietary data is shared with third parties,and the built-in mechanisms for logging and access control remain intact.Regularly update and audit security protocols.Workforce Impactand Change ManagementChallenge:Introducing AI can disrupt existing workflows and impact the workforce,leading t

99、o resistance.Solution:Engage with stakeholders and users early,provide training and support,and clearly communicate the benefits and changes and how they can resolve their pain points.Develop strategies for workforce transition and upskilling.drawings,equipment logs,maintenance records,3D models,ima

100、ges,and more).Continuous Improvement and MaintenanceChallenge:AI models require ongoing updates and maintenance to remain effective and relevant.Solution:Establish processes for continuous monitoring,performance evaluation,and retraining of AI models.Allocate resources for ongoing maintenance and im

101、provement.Addressing these challenges requires a strategic approach,involving cross-functional collaboration,investment in technology and skills,and a focus on long-term benefits.Fortunately,we have just the thingCognites Customer Success Frameworkto help you navigate these complexities successfully

102、(check out the Tools section at the end of this guide for more details).Chapter 1:Industrial Agents3637The Applicationsand Benefitsof Industrial AgentsIndustrial agents offer significant advantages over the more generic Generative AI approaches most use today,particularly in the context of asset-hea

103、vy industries.These advantages stem from their specialized design,precision,and integration capabilities tailored to the specific needs and challenges of industrial environments.Unlike broader LLMs,which provide general information and answers,industrial agents are equipped with a deep understanding

104、 of industry-specific processes and requirements that enables them to offer highly specialized solutions.This specialization ensures that the recommendations and optimizations provided are directly applicable and beneficial to the specific industrial context.While LLMs can process vast amounts of pr

105、e-existing information,industrial AI agents continuously monitor and analyze live data from industrial operations.This capability allows for immediate responses to changing conditions,proactive maintenance,and optimized operational efficiency.Additionally,industrial AI agents are tailored for predic

106、tive analytics specific to industrial applications.Their algorithms are fine-tuned to predict equipment failures,optimize maintenance schedules,and manage risks associated with complex machinery and processes.This precision in predictive capabilities is essential for minimizing downtime and enhancin

107、g the reliability and safety of operations,something broader LLMs are not equipped to handle with the same level of accuracy and relevance.Another key advantage is the integration of industrial AI agents with existing systems and infrastructure.Industrial AI agents are designed to work seamlessly wi

108、th industry-specific software,hardware,and operational protocols.This integration ensures minimal disruption and allows for smoother implementation and higher compatibility.Broader LLM searches,on the other hand,often require extensive customization to fit into industrial systems,which can be time-c

109、onsuming and less efficient.Furthermore,industrial AI agents are developed with a focus on compliance and quality control.They are equipped to monitor production standards,ensure regulatory compliance,and maintain product quality,which are critical aspects of asset-heavy industries.This level of con

110、trol and assurance is beyond the scope of what broader LLM searches can offer,as LLMs are primarily geared towards providing general information rather than enforcing industry-specific standards.The applications of targeted,real-time,highly integrated industrial agent solutions that can improve the

111、efficiency,safety,and reliability of energy,process manufacturing,and other industrial companies are seemingly limitless.A few examples include:Industrial agents represent a significant advancement in industrial AI applications,offering tailored,efficient,and practical solutions that address the uni

112、que challenges and needs of various industries.Their ability to enhance productivity,safety,and decision-making makes them invaluable in heavy-asset industries.Chapter 1:Industrial Agents3839AgentOrchestration and Agent EcosystemsAgent orchestration and agent ecosystems are transformative concepts i

113、n industrial environments,offering significant benefits in terms of efficiency,flexibility,and resilience.However,they also present challenges,particularly around complexity,integration,and security.Successfully implementing these concepts requires careful planning,robust infrastructure,and ongoing

114、management to ensure that agents can work together effectively and adapt to changing industrial needs.To level-set on some basic definitions,agent orchestration refers to the coordination and management of multiple autonomous agents to work together effectively.This involves ensuring that various ag

115、ents,each with specialized functions,collaborate seamlessly to achieve common goals such as optimizing production processes,maintaining equipment,and managing supply chains.Taking this one step further,agent ecosystems refer to the comprehensive network of interacting agents within a given organizat

116、ion.This includes not only the individual agents but also the infrastructure,protocols,and platforms that enable their interaction.In an industrial setting,an agent ecosystem encompasses all agents involved in production,maintenance,logistics,and more,working within a shared framework.An agent ecosy

117、stem is crucial in complex systems,particularly in industrial environments,where multiple autonomous agents must work together to achieve optimal performance.The agent ecosystem ensures that agents do not work in isolation,but rather in a coordinated manner,leading to increased efficiency.Orchestrat

118、ing multiple agents brings diverse perspectives and expertise to problem-solving,leading to more informed and effective decisions.The ecosystem leverages the strengths and capabilities of each agent to complement one another.By coordinating the activities of multiple agents,ecosystems can achieve re

119、sults that are greater than the sum of individual contributions.An agent ecosystem also allows for redundant systems that can take over tasks in case of failures,enhancing the overall resilience of the operation.AI Agents are dynamic and,thus,can adapt to changing circumstances,which more static Rob

120、otic Process Automation and custom-built automation tools are not capable of doing.As such,in case of the failure of one or more agents,ecosystems can reroute tasks and redistribute workloads among other agents,ensuring continuity and minimizing downtime.As the number of agents increases,ecosystems

121、help manage this complexity,ensuring that new agents are integrated smoothly,the system remains balanced,and all agents are aligned towards common objectives.In this way,an agent ecosystem can provide a global view of the entire system,allowing for optimization across all processes and agents rather

122、 than focusing on isolated parts.This ability enables the integration of different operational aspects(e.g.,production,maintenance,logistics)into a cohesive whole,enhancing overall system performance.In a smart factory,for example,agents might control various aspects such as supply chain logistics,p

123、roduction line operations,and quality assurance.An agent ecosystem ensures these agents work together to optimize production schedules,minimize downtime,and ensure product quality.Agent orchestration and agent ecosystems are vital for maximizing the potential of autonomous agents in industrial envir

124、onments.It ensures that multiple agents can work together effectively to achieve superior outcomes compared to isolated agents.This holistic approach is essential for modern industrial systems aiming to leverage AI and automation to their fullest extent.Chapter 1:Industrial AgentsAgent orchestration

125、 focuses on the coordination,task allocation,monitoring and control,and integration of individual agents.(Remember Iron Man and his many suits?)On the other hand,agent ecosystems focus on interoperability,scalability,and resilience of all agents.4041Does RAGStill Matter?Spoiler:It does.While industr

126、ial agents give AI applications new capabilities,Retrieval Augmented Generation(RAG)is necessary to effectively solve hallucination and data-freshness problems.A scalable,trustworthy,and safe generative AI implementation needs an excellent RAG solution to feed relevant,trustworthy,and up-to-date inf

127、ormation into large language models.Some companies are training LLMs from scratch,which requires a large investment.However,there has also been an influx of innovative methods for efficiently fine-tuning models,such as Parameter-Efficient Fine-Tuning(PEFT)6,QLoRA7,and prompt tuning8,which have becom

128、e popular.These techniques allow companies to“train”models on their data without large investments in data sets or hardware for model training.In addition,context window size(the capacity for LLMs to comprehend and retain knowledge)has greatly increased:Google has reported that they are experimentin

129、g with a context window of 10 million tokens.Given this development,RAG is surely obsolete.Or is it?Although the new techniques open up new possibilities,RAG remains a key ingredient in an effective enterprise generative AI framework.There are several reasons why fine-tuning and prompt tuning dont r

130、eplace the need for a good RAG system:6.Source:https:/ 7.Source:https:/arxiv.org/abs/2305.143148.Source:https:/arxiv.org/abs/2104.086919.Source:https:/ in Source AttributionDetermining the origin of information from the fine-tuned LLM can be difficult,making it challenging to differentiate between f

131、abrications and factual data.3.Issues with Information Access ControlLacking mechanisms to manage which data is accessible by different users.4.Ensuring Data IntegrityCompiling large,accurate data sets demands thorough data verification to guarantee the information remains current and correct.5.Info

132、rmation OverloadLike humans overwhelmed by too much information,a large context window holds vast data,but LLMs might struggle to pinpoint whats relevant,increasing the risk of hallucination and inaccuracies.92.Incorporating New Information or Removing Outdated InformationThis necessitates retrainin

133、g or re-fine-tuning,translating to additional time and financial investment.Chapter 1:Industrial Agents424310.Source:https:/ with a very large context window are a valuable addition to the toolbox,but its not a universal remedy.Although these models open up a variety of solutions,they have drawbacks

134、:1.PerformanceA long context window requires additional processing.This results in a time-to-first token(TTFT is the measurement of how long it takes before the LLM starts“typing”its response)that can be minutes.This is unacceptable for many,even most,use cases.2.ScalabilityWith additional processin

135、g needs comes reduced scalability.LLM providers operate with quotas.Google allows two requests per minute for Gemini Pro 1.5(in preview).Anthropic operates with a limit of 10,000,000 tokens per day and 40,000,000 tokens per minute for its top-tier subscription.Although there are always exceptions to

136、 such quota restrictions for large enterprises,the numbers are a clear testament to the fact that inference capacity is scarce.3.CostA natural consequence of reduced performance and scalability is increased cost.At the time of writing this,a single request to Anthropic Claude 3 with 200,000 tokens c

137、osts about$3.00 USD.This might be an acceptable price for some use cases but,for automated processing pipelines triggering hundreds or thousands of requests on a daily or hourly schedule,the cost will quickly explode.4.Environmental ImpactTraining LLMs is an extremely power-intensive operation.Altho

138、ugh energy efficiency is improving due to new hardware and algorithm improvements,interference remains a compute-intensive operation and will remain so for the foreseeable future.A recent report estimates that energy consumption associated with AI is expected to reach 0.5%of global electricity consu

139、mption by 2027.10A scalable,trustworthy,and safe generative AI implementation needs an excellent RAG solution to feed relevant,trustworthy,and up-to-date information to the LLMs.The Cognite contextualization engine and industrial knowledge graph allow accurate information to be retrieved.The context

140、ualization engine ensures data is connected across source systems,and the new AI service for populating structured knowledge graphs from unstructured documents ensures the industrial knowledge graph is as complete,accurate,and up-to-date as possible.Additional services for semantic search enable fil

141、tering information based on meaning rather than free-text search.This technique can even find information across multiple languages.These services make it easy to provide the best possible information to the LLMs,reducing the risk of hallucination while keeping the number of tokens required to proce

142、ss low.The strategies can be integrated with fine-tuning to achieve an optimal balance of timely,relevant updates and a deep understanding of the data.This hybrid approach allows fine-tuning to adapt the LLM to the general data landscape while RAG ensures the provision of current,relevant informatio

143、n,thereby yielding trustworthy outputs with traceable data sources.The key is building a representative benchmarking dataset for any generative AI capability to ensure consistent accuracy and performance across various use cases.Such data sets can also be used to compare the results from different L

144、LMs and assess how the different methodologies mentioned above perform.More to come on fine-tuning and benchmarking language models in the coming chapters.To conclude this section,while advancements in fine-tuning and training LLMs offer impressive capabilities,the integration of RAG remains indispe

145、nsable for delivering accurate and reliable AI-driven solutions.By combining innovative fine-tuning techniques with robust RAG systems from Cognite,companies can ensure their AI implementations are powerful and practical.Such a balanced approach enhances the efficacy of generative AI and maintains t

146、rustworthiness and relevance in industrial applications.Chapter 1:Industrial Agents4445From RAGto CAGRAG introduced the concept of enhancing generative models with retrieved information from external databases,significantly improving accuracy and relevance.However,RAG is limited by the dependency on

147、 the quality and relevance of retrieved documents.Yes,it is still important,but the next evolutionary step is Context Augmented Generation(CAG).CAG represents a more advanced approach,where the focus shifts from merely retrieving relevant documents to deeply integrating context from multiple sources

148、,including real-time data,sensor inputs,user interactions,and historical data.Also known as GraphRAG,this approach provides a richer and more dynamic context,enabling AI systems to generate more sophisticated and context-aware responses.CAG is a kind of contextual synthesizer,adapting data and infor

149、mation from various sources within the coherent and comprehensive context of a specific situation.Using the synthesized context,the generative model produces responses that are highly relevant and tailored to the specific needs and conditions of the environment.Key advantages to CAG include:Richer c

150、ontextual awareness:CAG leverages a broader and more dynamic set of context sources,leading to more nuanced and contextually appropriate responses.Real-time adaptation:By incorporating real-time data and user interactions,CAG can adapt its responses to changing conditions and requirements more effec

151、tively.Enhanced decision support:The comprehensive context provided by CAG improves the decision-making capabilities of AI systems,making them more useful in complex and dynamic environments.This does not mean CAG is without its difficulties.Integrating and synthesizing context from multiple diverse

152、 sources can be technically challenging.A robust,scalable,and flexible Industrial DataOps platform can play a significant role in overcoming this data integration problem by delivering:A unified data platform with centralized data management and an industrial knowledge graph to simplify data managem

153、ent and reduce the complexity of handling disparate types of data sources(e.g.sensor data,work orders,engineering drawings and 3D models and more).Data contextualization and metadata enrichment to link data points across systems and make it easier for CAG systems to interpret and utilize the data.Re

154、al-time data processing and streaming data support to ensure that the context used in generation is current and reflective of the latest conditions and changes in the industrial environment.Scalability to ensure that,as the number of data sources and the volume of data grow,the platform can continue

155、 to integrate and process data efficiently.Interoperability to ensure that it can work seamlessly with other systems and tools used in industrial environments.An example:A RAG system in a manufacturing environment might retrieve relevant maintenance manuals or past maintenance records to assist in t

156、roubleshooting equipment issues.A CAG system would go further by integrating real-time sensor data from the equipment,historical performance data,and current production schedules to provide a comprehensive diagnostic and maintenance plan.RAG SystemCAG SystemThe evolution from RAG to CAG marks a sign

157、ificant advancement in AI capabilities,moving from a model that relies on static document retrieval to one that dynamically integrates a rich and diverse set of contextual information.This evolution enables AI systems to generate more accurate,relevant,and contextually aware responses,making them mo

158、re effective in complex and dynamic environments.Chapter 1:Industrial Agents4647Large,Small,and Custom Language ModelsChapter 2Section 1:Making Generative AI Work for Industry48Introduction Large,Small,and Custom Language ModelsChapter 2:Large,Small,and Custom Language ModelsLanguage models are one

159、of the main ingredients in developing AI agents,providing them with the capability to understand and generate human-like text.Different models have distinct strengths,such as handling complex language structures,specific domains,or contextual understanding.To design effective AI agents,it is crucial

160、 to carefully select and leverage the appropriate language model,ensuring its capabilities align with the intended use and objectives of the agent.5051Understandingthe DifferenceChapter 2:Large,Small,and Custom Language ModelsLanguage modelslike GPT3.5,Claude,and Geminicome in various sizes and conf

161、igurations,each suited to different types of tasks and requirements.The primary differences between large,small,and custom language models revolve around their architecture,capabilities,resource requirements,and application suitability.In short,large models are versatile and suitable for a wide rang

162、e of tasks across various domains,small models are optimal for lightweight,specific applications with limited resources,and custom models excel in specialized applications where domain-specific knowledge is paramount.You must choose the right model for the right task.To make it easier to understand

163、the difference,weve put together a LM Cheat Sheet.53LLMsand Their Applicationin OperationsThe practical applications of LLMs in business operations are vast.LLMs can process vast amounts of text documents,extract relevant information,and summarize key findings.Increasingly,they are used to interpret

164、 engineering diagrams,designs,P&IDs,imagery,video recordings,voice commands,operational audio,vibration data,and other multi-modalities,all of which helps to extract insights from large volumes of unstructured data.For example,an LLM-based system can analyze maintenance reports,sensor logs,and opera

165、tor notes to help operators efficiently navigate and discover relevant data,leading to better decision-making and improved operational efficiency.LLMs can also play a vital role in industrial data analysis by assisting in critical activities such as anomaly detection and quality control.By ingesting

166、 historical data,sensor readings,and operational parameters,LLMs can learn to identify early signs of equipment failure,detect deviations from normal operating conditions,or pinpoint potential quality issues,supporting proactive maintenance strategies.LLMs are a powerful tool for industry,improving

167、operations in various ways that minimize downtime,reduce costs,and achieve higher overall efficiencies.With their ease of use,adaptability,and practical applications,LLMs offer a user-friendly solution that can streamline operations,automate tasks,gain valuable insights,and drive innovation in their

168、 respective industries.Chapter 2:Large,Small,and Custom Language Models5455So WhyDo We Need SLMs?SLMs,as the name implies,are a subset of machine learning models intentionally kept small in terms of their parameters and computational requirements.SLMs require less data and computational power to tra

169、in and can perform inference even faster than LLMs,which is critical for real-time applications.Additionally,SLMs are often fine-tuned for specific tasks or domains,which enhances their performance in those areas despite their smaller size.In general,SLMs are easier to maintain,update,and deploy acr

170、oss various systems and,due to their lower computational requirements,consume less energy,leading to lower long-term operational costs.Chapter 2:Large,Small,and Custom Language ModelsMore specifically,SLMs provide numerous benefits to industrial organizations,including:Enhanced natural language inte

171、rfaces:SLMs enhance voice-activated controls and natural language queries for industrial control systems,allowing operators to interact with machines and data systems more intuitively and efficiently.Knowledge extraction:SLMs can extract valuable insights from technical manuals,regulatory documents,

172、and historical records,providing decision-makers with critical information that supports informed decision-making.Real-time monitoring and diagnostics:SLMs can analyze operational logs and maintenance records in real-time to detect anomalies and predict equipment failures,helping minimize downtime a

173、nd optimize maintenance schedules.Predictive analytics:SLMs can process textual data,such as equipment logs and technician notes,to forecast potential issues before they escalate.Edge computing:SLMs can run on edge devices,enabling local data processing and decision making without relying on constan

174、t internet access,which is critical in heavy-asset industries in which connectivity to central servers might be limited.SLMs are crucial in heavy-asset industries due to their efficiency,cost-effectiveness,and ability to operate in resource-constrained environments.They enable real-time monitoring,e

175、nhance decision-making,optimize resource use,and ensure safety and compliance,making them indispensable tools for modern industrial applications.5657Custom Language Models in Industry Custom language models are tailored AI models designed to perform specific tasks or cater to particular domains by b

176、eing trained or fine-tuned on relevantoften proprietarydatasets.Custom models also incorporate nuances and specific terminologies of the targeted domain,enhancing their understanding and performance in that area.Key benefits of custom models include:Domain-specific accuracy:Custom models provide hig

177、her accuracy and relevance in industrial applications by understanding and processing domain-specific language and context.For example,in the energy sector,a custom model can interpret and generate reports,safety guidelines,and operational procedures with higher precision.Chapter 2:Large,Small,and C

178、ustom Language Models Efficiency and productivity:By automating complex tasks that require specialized knowledge,custom language models can significantly improve efficiency and productivity.In manufacturing,a custom model can optimize maintenance schedules,predict equipment failures,and streamline p

179、roduction processes.Enhanced decision-making:Custom models aid in better decision-making by providing insights and recommendations based on domain-specific data.For instance,in chemical processing,a custom model can analyze reaction outcomes and suggest optimal conditions,thereby improving yield and

180、 reducing waste.Regulatory compliance:In industries with strict regulatory requirements,such as pharmaceuticals or finance,custom models ensure that communications and documentations comply with industry standards and regulations,thereby minimizing legal risks.Cost reduction:By automating routine an

181、d complex tasks,custom language models reduce operational costs.For example,in supply chain management,a custom model can optimize logistics,reduce inventory costs,and improve demand forecasting.The best language model for industrial agents depends on various factors,including the specific requireme

182、nts,constraints,and goals of the industrial application.Each type of language modellarge,small,or customoffers distinct advantages and is suitable for different scenarios within industrial environments.5859Evaluating Large Language Models:Usefulness CorrectnessEvaluating LLMs involves assessing thei

183、r performance on specific tasks to understand how well they meet the intended use case requirements.As these models become increasingly integrated into various applications,ensuring their reliability,trustworthiness,and effectiveness is paramount.Without evaluations,you simply cannot know whether yo

184、ur LLM-based solutionwhether it is prompt engineering,Retrieval Augmented Generation(RAG),or fine-tuningis actually working,and neither can you improve it.Evaluation StrategiesEvaluations are a set of measurements used to check how well a model performs a task.An evaluation consists of two main comp

185、onents:benchmark data and metrics.While many benchmark data sets are available for LLMs,specialized tasks often require tailored data sets.For instance,if you want to use LLMs to generate a request body for your API service,you will need a dataset that includes examples of request bodies commonly us

186、ed in your application domain.Evaluation metrics are used to quantify the models performance on the benchmark data set.These metrics can broadly be classified into two main groups:traditional and nontraditional.Traditional metrics focus on the order of words and phrases,given a reference text(ground

187、 truth)for comparison.Examples include exact string matching,string edit distance,BLEU,and ROUGE.Nontraditional metrics leverage language models ability to evaluate generated text.Examples include embedding-based methods such as BERTScore,and LLM-assisted methods such as G-Eval,in which a powerful L

188、LM is instructed to evaluate the generated text.Methods where the performance of the system is estimated using pre-defined data sets,like the ones described above,are called offline evaluation.Another example of this is“sample testing”,whereby answers are randomly selected and checked against a sepa

189、rate(more expensive)LLM or by a human reviewer(much more expensive).This is similar to how quality control is done in many manufacturing spaces.While these offline evaluation methods are essential for ensuring that an LLM-based product feature has acceptable performance before deploying to users,the

190、y have their limitations and are usually not enough.Creating high-quality benchmark data sets takes time,and your data set can get outdated after releasing a feature and no longer represent the type of tasks users ask about.In addition,offline evaluation may not fully capture the complexity of real-

191、world user interactions.This is why offline evaluation must be complemented with online evaluations:the continuous evaluation of LLM-based features as they operate in a production environmentin real time,with real users and real data.After all,what matters most is whether the solution is actually us

192、eful for users in a production setting.All about evaluating LLMs:Overview of various evaluation methodsOnline evaluation includes user feedback,which can either be explicit,such as having users provide ratings like thumbs up or down,or implicit,such as monitoring user engagement metrics,click-throug

193、h rates,or other user behavior patterns.Note that both forms of feedback explicit and implicit are important.Explicit feedback is generally more accurate and less noisy than implicit feedback but tends to be much less abundant.To summarize,offline evaluations help you decide whether your LLM-based p

194、roduct feature has the minimum acceptable performance before deploying to users,while online evaluations are needed to ensure that your product continues to perform well in real-time user interactions,allowing you to monitor and improve its functionality over time based on live user feedback and beh

195、avior.So,what does this look like in practice?Chapter 2:Large,Small,and Custom Language Models6061Example use case:Natural language query to GraphQLNatural Language question to GraphQLOne of the AI features included in Cognite Data Fusion is the ability to search for data using natural language.Desc

196、ribed at a high level,the user input or question is converted into a(syntactically correct)GraphQL query,which is then executed(using the users credentials)to retrieve data from Cognite Data Fusion.The conversion from natural language to GraphQL is done through a set of promptsinstructions for large

197、 language models to generate a responsewhere each analyzes different aspects of the question and returns specific components of the GraphQL query.For instance,one prompt is designed to propose a suitable query operation(e.g.,get,list,or aggregate),another is designed to generate a suitable filter,an

198、d so on.Each prompt also has a corresponding post-processing step to ensure the validity of the generated output.For instance,a simple post-processing example is to check that the suggested query operation is a valid GraphQL query method.The outputs from all prompts are combined and used to construc

199、t a valid GraphQL query programmatically.Benchmark Data SetTo evaluate the previously described feature,multiple specialized data sets were curated with industry-relevant question-answers pairs.Our benchmark data set includes around ten different Cognite Data Fusion data models from energy and manuf

200、acturing sectors.Each model comprises tens to hundreds of real-life question-answer pairs,allowing evaluation across diverse scenarios.Below is an example test case from an Asset Performance Management(APM)data model:question:different formulations of a relevant question,that all can be addressed us

201、ing the same GraphQL query.relevantTypes:List of GraphQL types that are relevant to the question.queryType:The relevant GraphQL query and which type it should be applied on.queryType:Relevant filter.properties:Properties that are most relevant and should be returned given the context of the question

202、.summary:A short description of the suggested GraphQL query.Chapter 2:Large,Small,and Custom Language Models6263Evaluation MetricsThe prompts are evaluated using a mix of traditional and nontraditional methods.Simple string matching works for most fields but,for certain ones,like relevant properties

203、,we calculate standard metrics(recall,precision,and F1 scores)by comparing suggested properties to the ones in the benchmark dataset.The summary field is evaluated with an LLM-assisted approach,where a powerful language model(GPT-4)grades the suggested summary,given the ground truth summary.Command-

204、Line InterfaceFurthermore,we created a suite of developer tools to support a rapid feedback cycle while developing prompts and refining post-processing techniques.This suite comprises a Command-Line Interface(CLI)for assessing one or more prompts against a dataset or various datasets.It also include

205、s a Continuous Integration(CI)system that intelligently identifies the necessary evaluations to perform in response to modifications in a pull request(PR).It then compiles a report directly on the PR page,offering complete insight into the impact of the changes on the evaluation metrics.To monitor t

206、he models performance,especially for changes in the underlying base model,we calculate evaluation metrics daily across all available datasets.These metrics are tracked in Mixpanel.Pull RequestMixpanelMonitoring of offline evaluationmetrics in Mixpanel.CLI tool for evaluating promptson evaluation dat

207、asets in Mixpanel.Report of evaluation metricsin a Pull Request.Chapter 2:Large,Small,and Custom Language Models6465Online EvaluationEvaluating a proposed GraphQL query for syntactical accuracy is relatively straightforward.However,determining its semantic correctness and usefulness is significantly

208、 much harder.In this context,a semantically correct query retrieves relevant data based on user input or inquiry.To accurately gauge this or,more precisely,to obtain an indication of usefulness,online evaluation strategies are necessary.A few example metrics that are collected for this particular us

209、e case include:Thumbs-up and thumbs-down ratings:Users can provide feedback through simple thumbs-up or thumbs-down ratings,indicating their satisfaction or dissatisfaction with the retrieved data.User modifications of suggested filters in the UI:Users might adjust the filters suggested by the syste

210、m,tailoring the query to their specific requirements.These modifications may reflect that the suggested filter did not meet their expectations.User modification of the list of properties to display:Users might adjust the list of properties displayed in the result overview.These modifications may ref

211、lect that the properties shown initially did not meet their expectations.In addition,several other performance and utilization metrics related to latency,error responses,and wasted utilization of the LLMdue to service errors or any other unactionable responseare also collected.The online metrics ser

212、ve as a foundation for evaluating the effect of modifications to the feature through A/B testing.In such tests,various iterations of the prompt sequences are distributed across distinct user groups to determine the most effective version.Variations may include using different LLMs,alternative prompt

213、s,or varied pre-and post-processing approaches.Moreover,these online metrics are instrumental in establishing Service Level Objectives(SLOs)for response times and end-user error rates,as well as user satisfaction measured via the satisfaction metrics mentioned earlier in this section.Usefulness Corr

214、ectnessIts important to emphasize that usefulness differs significantly from correctness.While correctness can be measured explicitly and objectively,usefulness is inherently subjective.Hence,the aforementioned metrics serve as indicators of whether users find the tool valuable or not.These metrics,

215、derived from user interactions and feedback,offer valuable insights into the practical utility of the GraphQL query and help refine the system for optimal user satisfaction.Chapter 2:Large,Small,and Custom Language Models6667Choosingthe Right Model Using AutoLLMAs we mentioned before,the best langua

216、ge model for industrial agents depends on various factors,including the specific requirements,constraints,and goals of the industrial application.Each type of language modellarge,small,or customoffers distinct advantages and is suitable for different scenarios within industrial environments.So how d

217、o you choose the right model for your needs if you are not an AI engineer?AutoLLM is an automated process designed to help users select the most suitable language model for their specific needs.This process leverages various criteria,including performance,cost,and specific use case requirements,to r

218、ecommend the best model.AutoLLM refers to automated systems and frameworks designed to select,configure,and optimize language models for specific tasks or applications.It leverages various criteria,including performance,cost,and specific use case requirements to simplify the complex process of choos

219、ing the most suitable model,tailoring it to the users needs,and ensuring optimal performance.With autoLLM,a user provides details about their specific use case,including the type of task(e.g.,text generation,summarization,classification),performance expectations,resource constraints and budget,and t

220、he system uses algorithms to match user requirements with the most suitable models from the database.For example,models that analyze images and sensor data from production lines are better suited for quality control and defect detection,while models that analyze geological and production data are be

221、st for optimizing reservoir management strategies.AutoLLM also automates the configuration and tuning of the recommended model as needed to optimize model performance for the specific task.Once the optimal model is selected and configured,AutoLLM can automate the deployment process,integrating the m

222、odel into the users application or workflow.Post-deployment,the system monitors model performance in real-time,providing feedback and making adjustments as necessary to maintain optimal performance.AutoLLM simplifies the process of choosing and deploying the right LLM,making advanced AI accessible t

223、o users without deep machine-learning expertise.Not only does the system save time and resources,but it also continuously refines and adjusts models to ensure sustained performance improvements over time,ensuring better performance and more accurate results.Additionally,autoLLM helps reduce unnecess

224、ary resource consumption and associated costs by balancing performance with computational efficiency and tailoring solutions to fit within specified budget constraints,optimizing for cost-effectiveness.By automating the complex tasks of model selection,configuration,and optimization,AutoLLM empowers

225、 users to leverage advanced AI capabilities without requiring deep technical expertise.Thus,autoLLM makes industrial agents more accessible,efficient,and effective.And it is these agents that will enhance predictive maintenance,knowledge management,quality control,supply chain optimization,process a

226、utomation and more,leading to improved efficiency,cost savings,and better decision-making.Chapter 2:Large,Small,and Custom Language Models6869Performance BenchmarkingPerformance benchmarking for language models is the process of evaluating and comparing their efficiency,accuracy,and overall effectiv

227、eness in performing specific tasks.This process involves systematically measuring various performance metrics to ensure that the models meet the desired standards and are suitable for their intended applications.Performance benchmarking involves assessing language models against predefined metrics a

228、nd standards.This evaluation helps in understanding the strengths and weaknesses of the models,guiding improvements,and ensuring that they are fit for purpose.Task-specific metrics:Metrics such as accuracy,precision,recall,F1-score,and BLEU score,which are relevant to specific tasks like classificat

229、ion,translation,summarization,etc.Resource metrics:Measures of computational efficiency,such as inference time,memory usage,and processing speed.Scalability metrics:Metrics to assess how well the model handles increasing data volumes and user demands.Benchmark suites:Standardized sets of tasks and d

230、atasets are used to test models uniformly.Benchmarking provides a quantitative basis for comparing models,ensuring that decisions are based on objective data rather than subjective judgments.It helps in identifying the most efficient models that offer the best performance relative to resource consum

231、ption.The benchmarking process involves comparing the performance of multiple language models across a variety of standardized tasks and datasets,aiming to provide insights into how different models perform relative to each other,highlighting their strengths,weaknesses,and overall suitability for va

232、rious natural language processing(NLP)tasks.We use benchmarking to determine the suitability of the model for a specific task.For example,evaluating a models accuracy and inference time in generating text summaries and providing insights into its task-specific strengths and weaknesses.Benchmarking m

233、odels against each other provides an objective basis for comparing the strengths and weaknesses of different models,offers insights into which models are more suitable for specific tasks based on empirical data,and helps stakeholders make informed decisions about model selection and deployment strat

234、egies.Chapter 2:Large,Small,and Custom Language Models70Semantic Knowledge GraphsChapter 3Section 1:Making Generative AI Work for Industry72Introduction SemanticKnowledge GraphChapter 3:Semantic Knowledge GraphKnowledge graphs are crucial to AI agents as they provide structured,interconnected inform

235、ation that enhances the agents ability to understand and process complex relationships between data.By leveraging knowledge graphs,AI agents can deliver more accurate,context-aware insights and recommendations,improving their effectiveness in tasks such as data analysis,decision-making,and problem-s

236、olving.Integrating knowledge graphs ensures AI agents can access and utilize rich,domain-specific knowledge,elevating their overall performance and utility.7475Definingthe Semantic Knowledge GraphChapter 3:Semantic Knowledge GraphKnowledge graphs map data relationships,capture interconnections,and t

237、race data lifecycles.They are constructed by combining data sets from diverse sources,each varying in structure.The harmony between schemas,identities,and context contributes to the coherence of this comprehensive data repository.Schemas establish the fundamental framework upon which the knowledge g

238、raph is built.Identities efficiently categorize the underlying nodes.Context determines the specific setting in which each piece of knowledge thrives within the graph.Knowledge graphs use machine learning to construct a holistic representation of nodes,edges,and labels through a process known as sem

239、antic enrichment.Knowledge graphs can discern individual objects and comprehend their relationships by applying this process during data ingestion.This accumulated knowledge is then compared and fused with other data sets that share relevance and similarity.You have likely heard us reference the ind

240、ustrial knowledge graph before.This open,flexible,labeled property graph represents your operations by focusing on the intricacies and specifics of industrial processes,machinery,operations,and related data.It liberates the data that has been locked in different systems and applications(high-frequen

241、cy time-series sensor data,knowledge hidden in documents,visual data streams,and even 3D and engineering data)and makes it meaningful and manageable.A semantic knowledge graph captures the relationships between entities in a way that humans and machines understand.It enables question-answering and s

242、earch systems to retrieve comprehensive responses to specific queries.Semantic knowledge graphs are time-saving tools that streamline manual data collection and integration efforts to bolster decision-making processes.The powerful combination of industrial and semantic knowledge graphs delivers a st

243、ructured representation of information that allows industrial organizations to understand the relationships and connections within complex datasets better.These data relationships are made possible through contextualization pipelines that help create and maintain a dynamic knowledge graph;thus,addre

244、ssing three key challenges:Overcoming data silos:In industrial settings,data often resides in numerous silos,leading to duplication and ambiguity in meaning.Knowledge graphs play a pivotal role in breaking down these silos,providing a unified view of data,and improving the understanding of usage and

245、 consumption patterns.Unleashing unstructured data:By employing standardized metadata,the knowledge graphs allow for the categorization and management of information,enhancing the utilization of unstructured data present in documents,images,and videos(another common data silo)and turning that data i

246、nto actionable insights.Enhancing business insights:The explicit contextualized knowledge,rules,and semantics embedded in knowledge graphs empower AI applications to provide high-quality,trusted insights that are absolutely necessary for working in the industrial domain.They also allow subject matte

247、r experts to make high-quality decisions,enhancing business processes,workflows,and operations.However,a knowledge graph is only as valuable as the data it can access.7677Knowledge Graphs and DataA knowledge graph enables an industrial organization to extract value from unstructured and siloed data

248、sources.Establishing a dynamic and interoperable industrial knowledge graph with access to high-quality contextualized data must be the first step for any organization that wants to implement generative AI initiatives that improve operations and accelerate time to value.In this example,the left diag

249、ram above illustrates a simplified version of an industrial knowledge graph of a centrifugal pump.Depending on the persona,users may explore a problem with the pump from multiple entry points.Maintenance may start with the latest maintenance report,an operator may use the time series,and a remote SM

250、E may begin with the engineering diagram(e.g.,P&ID).The maintenance report,the work order,the time series values,and the engineering diagrams are each in separate systems.Having all this data connected in the industrial knowledge graph creates a seamless experience,regardless of the starting point.T

251、his simple example illustrates the importance of data contextualization across different systems.Cognites AI-powered data contextualization capabilities power the industrial knowledge graph(as seen on the right side)to provide access to the maintenance report,work order,time series,and more in a sin

252、gle location.With the industrial knowledge graph as the foundation,data is understood and structured to meet the specific needs of users or use cases,making it easier for all stakeholders to find the necessary information,and view and understand relationships between data objects.Simple aggregation

253、of digitized industrial data is a significant step forward from the silos and inaccessibility that often plague large enterprises.However,to provide simple access to complex data,the variety of industrial data types must be accounted for,and the semantic relationships that drive scalable utilization

254、 of this data must be incorporated to support interactive user experiences.Codifying this context as an industrial knowledge graph is vital to enabling consistent,deterministic navigation of these meaningful relationships.Chapter 3:Semantic Knowledge Graph7879Lets give an example.An asset hierarchy

255、is ideal for addressing use cases related to asset performance management(APM).The relationships resource type also allows the organization of assets in other structures besides the standard hierarchical asset structure.Instead,assets can be organized by their physical location,where the grouping no

256、des in the hierarchy are buildings and floors rather than systems and functions.Building a model with this structure opens the possibility of solving new use cases like product traceability,where the physical connections of the assets through the production process must be known.The knowledge graph

257、enables you to capture various data perspectivesChapter 3:Semantic Knowledge GraphIn this way.data becomes an asset,with reusable analytics and scalable models,shareable across many users.This industrial knowledge graph encourages data reuse by creating a user-friendly architecture.By leveraging dat

258、a effectively and rapidly,the organization can address business opportunities quickly and at scale.8081Knowledge Graphs and Generative AI According to Gartners Emerging Tech Impact Radar,Generative AI report,knowledge graph adoption has rapidly accelerated with the growing use of AI because knowledg

259、e graphs provide the explicit knowledge,rules,and semantics needed in conjunction with AI/ML methods for pattern recognition.In other words,knowledge graphs deliver trusted and verified data to LLMs and provide rules to contain the model.The industrial knowledge graph is the foundation for a data mo

260、del and provides the point of access for data discovery and application development.The most prevalent application of data modeling is to unlock the potential of industrial digital twins.The advantage of data modeling for digital twins is to avoid the singular,monolithic digital twin expected to mee

261、t the needs of all and focus on creating smaller,tailored twins designed to meet the specific needs of different teams.The above graphic shows that a digital twin isnt a monolith but an ecosystem.What is needed is not a single digital twin that perfectly encapsulates all aspects of the physical real

262、ity it mirrors,but rather an evolving set of digital siblings who share a lot of the same DNA(data,tools,and practices)but are built for a specific purpose,can evolve on their own,and provide value in isolation.Like Metcalfes Law and the understanding of the exponential value of a network,data in th

263、e industrial knowledge graph becomes increasingly valuable as people use,leverage,and enrich that data.More useful and high-quality data leads to more trusted insights.More trusted insights lead to higher levels of adoption by subject matter experts,operations and maintenance,and data science teams.

264、A user-friendly,AI-powered experience ensures adoption and use will grow,and this cycle repeats exponentially.In this way,knowledge graphs are a key underlying technology and act as the backbone for generative AI solutions across business functions that will drive business impact,including:Digital w

265、orkplace(e.g.,collaboration,sharing and search)Automation(e.g.,ingestion of data from content to robotic process automation)Data exploration(e.g.,providing deeper insights into structured and unstructured data)Data management(e.g.,metadata management,data cataloging,and data fabric)Chapter 3:Semanti

266、c Knowledge Graph83Despite the undeniable benefits of knowledge graphs,Gartner identifies several challenges to successful implementation.Lets take a look at how we can address these challenges:As knowledge graphs transition from prototypes to production,methods to maintain their scalability while e

267、nsuring reliable performance,handling duplication,and preserving data quality are still evolving.Enabling internal data to interact with external knowledge graphs(meaning connecting data and graphs that vary in scope,ownership,data types,etc)seamlessly remains challenging.Overcoming this hurdle is v

268、ital for the creation of a truly interconnected and interoperable industrial ecosystem.Particularly among small and midsize businesses,expertise in knowledge graphs is scarce.Identifying and accessing third-party providers with the necessary proficiency becomes a significant obstacle.ChallengeImmatu

269、re Scaling MethodsChallengeInteroperabilityChallengeScarcity of In-House ExpertiseTo provide reliable performance and scalability,organizations must ensure that their knowledge graphs are powered by contextualization services to provide high-quality,trusted insights that lead to higher levels of ado

270、ption by the teams across the enterprise.To establish and maintain interconnection and interoperability,we must ensure access to fully documented and open APIs that help facilitate connections between different data systems,industry standards models,or third-party applications.Plus,strong contextual

271、ization capabilities ensure the necessary background for meaningful integration and interpretation of information,especially when that information is trapped in a siloed data source.Working with a third-party provider with expertise in building industrial knowledge graphs and industrial data managem

272、ent should not be that scary,especially if you know what to avoid during decision-making when purchasing software and what software deployment type works best for your organization and goals.SolutionSolutionSolutionChallenge 1Challenge 2Challenge 3Chapter 3:Semantic Knowledge Graph8485To be effectiv

273、e in complex industrial settings,a knowledge graph must include:Automated population with contextualization,cross-source-system IT,operational,and engineering data;A robust,well-documented API integration;and Extremely performant,real-time,flexible data modeling.Chapter 3:Semantic Knowledge Graph868

274、7Section 2The BusinessValue of AIThe Cognite Atlas AI Definitive Guide to Industrial AgentsChapter 4AI Is the DrivingForce for IndustrialTransformation.90 4.1 Verdantix View:Industrial DataOps in 2024.94 4.2 AI Will Deliver Untapped Value for Asset-Heavy Enterprises.102 4.3 Democratizing Data:Why AI

275、-Infused Industrial DataOps Matters to Each Data Stakeholder.104Chapter 5Use Cases.108 5.1 Industrial Use Cases Require a System of Engagement.110 5.2 Cognite Data Fusion:An SOE to Scale Operational Use Cases.114 5.3 Improving RCA with AI Agents and Industrial Canvas.122 5.4 Examples of Industrial A

276、I Agents.126AI Is the Driving Force for Industrial TransformationChapter 4Section 2:The Business Value of AI90IntroductionAI Is the Driving Force for Industrial TransformationChapter 4:AI is the Driving Force for Industrial TransformationThere are two discomforting truths within digital transformati

277、on across asset-heavy industries:Digitalization Proof-of-Concepts(PoCs)are commonplace.Real Return on Investment(ROI)isnt.Organizations invest billions in cloud data warehouses and data lakes.Most data ends there,unused by anyone for anything.The ChallengeOnly one in four organizations extracts valu

278、e from data to a significant extent.Data dispersion and a lack of tools and processes to connect,contextualize,and govern data stand in the way of digital transformation.The OpportunityIndustrial DataOps,infused with AI,promises to improve the time to value,quality,predictability,and scale of the op

279、erational data analytics life cycle.It provides the opportunity to offer data science liberation within any product experience while simultaneously allowing subject matter experts to acquire,liberate,and codify domain knowledge through an easily accessible and user-friendly interface.This is a stepp

280、ing stone to a new way of managing data in the broader organization,enabling it to cope with growing data diversity and serve a growing population of data users.Value Capture Reality:Despite supportive technology trends in abundance,data activation remains stuck in POC purgatory9293Verdantix View:In

281、dustrial DataOpsin 2024Joe Lamming Senior Analyst,Operational ExcellenceAs we enter the second half of 2024,the true potential of AI in industrial operations is becoming increasingly evident.Industrial DataOps continues to be foundational andin many waysat the forefront of this revolution,offering a

282、 robust framework for managing and effectively utilizing vast quantities of data generated by asset-heavy industries.At Verdantix,we define industrial data management solutions as software that facilitates a professional approach to managing data,improving data quality,and facilitating collaboration

283、 between domain experts and data scientistsall while constructing data pipelines for tracking and verifying data origins from the moment of acquisition through usage and to eventual deletion.The Evolution and Importanceof Industrial DataOpsVerdantix has been a keen observer of the data analytics and

284、 AI landscape since our inception in 2008.We have witnessed significant shifts,from the IoT boom in the 2010s to the rise of Big Data for industrial asset management.Our focus sharpened on Industrial DataOps around 2021 and today,the drivers for industrial data management solutions are stronger than

285、 ever:1.Successful AI Deployments:Effective AI models require high-quality,real-time data from diverse sources.Industrial DataOps ensures that this data is properly aggregated,cleaned,and ready for use in sophisticated AI models for anomaly detection,recommendations,and predictions.2.Top-Down Pressu

286、re:Business leaders need access to both granular and big-picture data to make informed decisions quickly.Industrial DataOps provides this data in a concise,visualized format,enabling self-service analytics and real-time decision-making.3.Operational Efficiency:Initiatives aimed at cost reduction,pro

287、duction optimization,and operational safety all benefit from robust data acquisition and seamless integration with analytics tools.Industrial DataOps plays a critical role in identifying improvement areas and boosting collaboration between departments.4.Sustainability Reporting:With increasing regul

288、atory pressure for ESG and sustainability reporting,industries need robust data management systems to comply with these requirements.Industrial DataOps ensures that all necessary data is accurately tracked and reported.Market Trends and ChallengesThe Verdantix Market Size&Forecast for industrial AI-

289、focused analytic solutions,published in December 2023,reveals a compound annual growth rate(CAGR)of just under 24%significant in the industrial software space.The biggest spend areas are asset condition monitoring,predictive maintenance,and product and process management,with substantial growth also

290、 seen in supply chain optimization and production and process management.Chapter 4:AI is the Driving Force for Industrial TransformationSurvey data from our Operational Excellence Global Corporate Survey also highlights the importance and challenges of data aggregation,access,collaboration,and conte

291、xtualization.Data aggregation remains a critical challenge for many,particularly in process manufacturing and energy sectors.Data access and collaboration between data scientists,operations,and maintenance executives are also seen as significant challenges,with a clear need for improvement in the en

292、ergy sector.9495Overcoming Data Integration ChallengesA major challenge in deploying AI in industrial settings is data integration.We see a number of vendors tackle this issue with platform solutions that provide comprehensive data modeling services able to orchestrate diverse data types from dispar

293、ate sources.We have witnessed Cognite Data Fusions ability to enable the creation of detailed data models that standardize information across equipment and processes,making it easier to analyze and use.Enhancing Data Qualityand ContextualizationEnsuring data quality is another critical aspect.Automa

294、ted monitoring of sensor data,cross-referencing tags,and enforcing data governance frameworks are essential for maintaining the high data quality needed for trustworthy analytics.Additionally,contextualizing data assembling it into digestible formats for data scientists and decision-makers expedites

295、 its utility and resultant time-to-value.Getting Access to DataIs Industrial Firms#1 PriorityProviding Low-Code Accessto Both Data Scientistsand Operations and Maintenanceis Top of MindEffective Data Contextualization Continues to Be an Important Challenge for Industrial FirmsManaging Data Quality T

296、hrough Robust Governance Is a Significant Priority for Firms in 2024Source:Verdantix Operational Excellence Global Corporate Surveys 2022 and 2023.Notes:Figures rounded to the nearest integer.Percentages lower than 5 are written as numbers.Chapter 4:AI is the Driving Force for Industrial Transformat

297、ion9796Data-Driven Decision SupportSuch as AI Analytics Are Expectedto See Nearly 24%CAGR Until 2028Industrial Firms Are Interestedin Generative AI Necessitating Deeply Contextualized,Accessible DataThe Impact of Generative AIAIs role in industrial transformation is multifaceted.One of its key appli

298、cations is in predictive maintenance,with ML-driven anomaly detection and forecasting models analyzing data from sensors to predict equipment failures before they occur,thereby providing the opportunity to optimize production around scheduled downtime and reduce maintenance costs.AI tools are also u

299、sed in asset condition monitoring,process optimization,and supply chain management,driving similar efficiency gains across operations.Generative AI is poised to fundamentally reshape core administrative tasks within industrial operations through vastly more capable automation and improved data disco

300、verability.Our survey data indicates that only 5%of respondents foresee no changes due to generative AI.Many expect Gen AI to automate administrative tasks,improve data access,and enable direct querying of technical documentation.However,challenges such as regulatory compliance and the complexity of

301、 deploying AI models,especially those fine-tuned on domain-specific data,must be addressed.Partnering with data management software vendors can help overcome these hurdles and achieve successful AI implementations.Source:Verdantix Market Size and Forecast Industrial AI Analytics 20222028(Global)Note

302、:Analysis published December 2023Source:Global Corporate Survey 2023:Operational Excellence Budgets,Priorities and Tech PreferencesChapter 4:AI is the Driving Force for Industrial Transformation9998Innovative Data Modeling StrategiesKey to making AI for industry work are graph databasesor industrial

303、 knowledge graphs.These databases capture and illustrate semantic relationships between entitiesworkers,assets,process documentation,and even 3D models.This approach reduces duplicate siloed data and enhances data discovery,improving search capabilities and enabling big-picture network analysis for

304、critical processes.Such semantic data models also provide a solid foundation for AI agents,enterprise search engines,and AI copilots,reducing the risk of hallucinations by guardrailing outputs with clear,contextualized data.Collaboration with data management software vendors is crucial to overcoming

305、 these hurdles and achieving successful AI implementation.The best data management solutions deliver six critical capabilities:Easy-to-configure data connectors Pre-built data quality management Integrations with popular BI tools Contextualization through both semantic and asset hierarchies Data dis

306、covery with knowledge graphs Collaborative GUIs augmented with Gen AIAs we look toward the close of 2024,it is evident that AI,supported by robust Industrial DataOps,will be the driving force behind industrial transformation.The integration of AI into data management processes not only enhances oper

307、ational efficiency,but also empowers industries to navigate the complexities of modern industrial operations with greater agility and insight.Industrial DataOps platform providers at the frontier are a critical pillar of this revolution,enabling industries to harness the full potential of AI and dat

308、a-driven decision-making.The future of industrial operations is not just about managing data;it is about transforming that data into information actionable intelligence that drives better decisions,better outcomes,and better explainability.The journey towards a data-driven future was already well un

309、derway.Now we have rocket fuel.Chapter 4:AI is the Driving Force for Industrial Transformation100101AI Will Deliver Untapped Valuefor Asset-Heavy EnterprisesTo extract the value of industrial data insights adequately,it is essential to make operationalizing data the core of your business strategy.Da

310、ta must be available,useful,and valuable in the industrial context.Operational teams need a robust data foundation with a strong data context and interpretability backbone,all while applying generative AI to accelerate workflows that optimize production and make operations more efficient.Efficient D

311、ata Managementand Improved Data AccessibilityAugmented Workflowsand Process Improvements,Driving Innovation at ScaleRapid Development of Use Casesand Application Enablement Enterprise Data Governanceas a By-Product and PersonalizedAI Tools1.2.3.4.A strong data foundation is required to remove the ri

312、sk of hallucinations and increase AI readiness.An Industrial DataOps foundation maximizes the productive time of data workers with automated data provisioning,management tools,and analytic workspaces to work with and use data safely and independently within specified governance boundaries.The approa

313、ch can be augmented with AI-based automation for various aspects of data managementincluding metadata management,unstructured data management,and data integrationenabling data workers to spend more time on use case development.Using Using generative AI-powered semantic search,what used to take your

314、process engineers,maintenance workers,and data scientists hours of precious time will take only a few seconds.With the guidance of generative AI copilots,users can generate summaries of documents and diagrams,perform no-code calculations on time series data,conduct a root cause analysis of equipment

315、,and more.Time spent gathering and understanding data goes from hours in traditional tools to seconds.Now,users can spend more time driving high-quality business decisions across production optimization,maintenance,safety,and sustainability.Too often,digital operation initiatives get trapped in PoC

316、purgatory,as scaling pilots takes too long or is too expensive.Using an AI-infused Industrial DataOps platform shortens the time to value from data by making PoCs quicker and cheaper to design and offering operationalizing and scaling tools.These copilot-based approaches leverage the power of natura

317、l language to understand and write code based on published API documentation and examples to support development processes.Generative AI further improves ML training sets of ML models by generating synthetic data,enhancing the data set used for training,enhancing process efficiency,and optimizing pr

318、oduction.Some common use cases in asset-heavy industries are maintenance workflow optimization,engineering scenario analysis,digitization of asset process and instrumentation diagrams(P&IDs)to make them interactive and shareable,and 3D digital twin models to support asset management.By having a stro

319、ng Industrial DataOps foundation,you can then empower users to adapt AI models to cater to their specific requirements and tasks,using generative AI to enhance data onboarding,complete with lineage,quality assurance,and governance,while a unique generative AI architecture enables deterministic respo

320、nses from a native copilot.Additionally,Industrial Canvas overcomes the challenges of other single pane-of-glass solutions,which often over-promise capabilities and are too rigid with prescribed workflows.This prevents users from working with the data how they choose by delivering the ultimate no-co

321、de experience within a free-form workspace to derive cross-data-source insights and drive high-quality production optimization,maintenance,safety,and sustainability decisions.If implemented successfully,an AI-augmented data platform provides consistency and ROI in technology,processes,and organizati

322、onal structures,with better operations data quality,integration and accessibility,and stewardship.It should also enhance data security,privacy,and compliance with tracking,auditing,masking,and sanitation tools.AI to enable rapid ingestion and contextualization of large amounts of data brings a parad

323、igm shift in how the organization accesses business-critical information,improving decision-making quality,reducing risk,and lowering the barriers to(and skills for)data innovation.Chapter 4:AI is the Driving Force for Industrial Transformation102103Democratizing Data:Why AI-Infused Industrial DataO

324、ps Matters to Each Data StakeholderExtracting maximum value from data relies on applying advanced models to produce insights that inform optimal decision-making,empowering operators to take action confidently.Turning insight into action is what we mean by operationalizing data into production for va

325、lue.But for every person who can speak code,hundreds cannot.Generative AI will change how data consumers interact with data.It facilitates a more collaborative working model,in which non-professional data users can perform data management tasks and develop advanced analytics independently within spe

326、cified governance boundaries.This democratization of data helps store process knowledge and maintain technical continuity so that new engineers can quickly understand,manage,and enrich existing models.It is about removing the coding and scripting and bringing the data consumption experience to the h

327、uman user level.Why It Matters to ExecutivesChapter 4:AI is the Driving Force for Industrial TransformationMaking the data speak human is the only way to address the Achilles heel of practically all data and analytics solutions,especially those for heavy-asset industry verticals.These organizations

328、face many challenges:an aging workforce,extreme data type and source system complexity,and very low classical data literacy among SMEs those needing data to inform their daily production optimization,maintenance,safety,and sustainability decisions.104105Solving the industrial data and AI problem is

329、critical to realizing value from digitalization efforts.Benefits can be measured from streamlined APM workflows,improved SME productivity,optimized maintenance programs,and real-time data efficiencies,such as:Productivity savings due to improved SME efficiency.Industrial DataOps provides data access

330、ibility and visibility,transforming how data scientists and SMEs collaborate.Reduced shutdown time.The opportunity cost of large industrial assets being out of production is significant.Using a digital twin and better component data visibility,SMEs are able to safely minimize shutdown periods when d

331、ata anomalies arise.Real-time data access enables Improvement in productivity.Live data access enhances operational flexibility and decision-making by increasing site safety,improving predictive maintenance,and raising machine performance.Optimized planned maintenance.Cognite Data Fusion creates con

332、textualized data to optimize planned maintenance by analyzing and interpreting available resources,workflows,and component life cycles.Energy efficiency savings.Intelligent data can be used to reduce energy use and,thus,operational costs.Optimization of heavy machinery and industrial processes.Healt

333、h and safety.Reduce the amount of human movement through potentially dangerous hot areas,reducing risks to employee health and safety.Environmental,social,and governance(ESG)reporting.Why It Matters to IT and Digital TeamsWhy It Matters to Domain ExpertsChapter 4:AI is the Driving Force for Industrial Transformation106107Use CasesChapter 5Section 2:The Business Value of AI108Industrial Use Cases R