使用 Amazon Bedrock 构建可用于生产环境的 AI 应用 [重复].pdf

1、 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.P E X 4 0 4-RBuilding Production-Ready AI Applications with Amazon BedrockMassimiliano AngelinoAI/ML Spec.Solutions ArchitectAmazon Web ServicesEmir AyarSr.Solutions

2、 ArchitectAmazon Web ServicesMatt Houghtonhe/himData&AI ArchitectCDL 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.The prototype to production“chasm”Excitement and potentialPOCMeaningful business value 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.2025,Amazo

3、n Web Services,Inc.or its affiliates.All rights reserved.Prepare for production 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.Understand failure modesMake sure systems behave the way they are designed for and are secureGenerative AI systems are non-deterministic,it is not possib

4、le to exhaustively test all possible outcomesInputs are unstructured and contain natural languageAI Agents have autonomy to plan and choose toolsGenAI systems can fail silently 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.O B S E R V A B I L I T YTrack what mattersGain visibili

5、ty into our generative AI systems in productionMonitor latency,tokens,API errors by model/use caseTrack agent loops,tool calls,auth requestsLog LLM calls with correlation IDsMonitor system metrics,like CPU,memoryCapture feedback from user:thumbs up/down,textual,implicit(based on behaviour)2025,Amazo

6、n Web Services,Inc.or its affiliates.All rights reserved.E V A L U A T E A N D I M P R O V EEvaluate&ImproveUnderstand the performance of your system through systematic testing Use all results good and bad to improve the quality of the system,craft reference datasetsBuild replay capabilitiesUse auto