大规模构建和管理对话式人工智能：来自 Alexa+ 的经验教训.pdf

1、 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.A M Z 3 0 1Building and Managing Conversational AI at Scale:Lessons from Alexa+Brittney HurstShe/HerAWS Gobal LeadAmazon Devices&ServicesLuu TranHe/HimEngineerAlexa

2、CoreSai RupanagudiHe/HimProductAlexa AI 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.AgendaIntroductionAbout Amazons AlexaAlexa+Design Considerations Solution and journeyKey Takeaways 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.2025,Amazon Web Services,In

3、c.or its affiliates.All rights reserved.2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.Customer ExpectationsThe Challenge:re:Invention without Disruption 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.2025,Amazon Web Services,Inc.or its affiliates.All rights r

4、eserved.Alexa+A step change 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.Design ConsiderationsModel Flexibility Latency DeterminismAccuracy 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.Accuracy 01Intent(Understanding)ActionsCompounding errors 2025,Amazon W

5、eb Services,Inc.or its affiliates.All rights reserved.Accuracy 02RoutingTools/Expert/AgentHard to recover if wrong 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.Accuracy 03APIsNameParametersValues 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.Accuracy 04Tuni

6、ng and refactoringExemplarsRefactoringSkip the obvious 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.LatencyTraditionalTechniquesParallelizationStreamingPrefetching01 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.LatencyTokensInput vs.Out