使用 Amazon Aurora 构建经济高效的基于 RAG 的生成 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.2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.D A T 4 2 8-RJim MlodgenskiShayon SanyalBuild a cost-effective RAG-based gen AI application with A

2、mazon AuroraSenior Principal EngineerAmazon RDSPrincipal Database Specialist SAAmazon RDS 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.AgendaEcoNovas GenAI JourneyModel choiceRetrieval Augmented Generation(RAG)optimizationsCaching and memory managementArchitectural and infrastr

3、ucture optimizations$2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.EcoNovas GenAI JourneyDeployed GenAI agents for optimizing operationsRapid adoption created unpredictable costs Latency issues hurt user experience 2025,Amazon Web Services,Inc.or its affiliates.All rights reserv

4、ed.Workload assumptions100K agent tasks/dayEach task averages:2K input tokens,1K output tokensInitial DB storage 1TB,growing 20GB daily on 2x db.r7i.xlarge instances$63,000$1,086$345$114$220,500$2,715$380$125$0$50,000$100,000$150,000$200,000$250,000LLM CostsDB InstancesDB StorageEmbeddingsMonthly co

5、sts(us-west-2)RAGAgentic RAGMain driver of costsNote:Pricing as of December 2025;subject to change 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.Model choice 2025,Amazon Web Services,Inc.or its affiliates.All rig

6、hts reserved.Model Distillation:75%cost reduction,500%faster inference,fixed tokensOverlap chunks 10-15%for context continuityMetadata filtering:40-60%search space reduction(category,date,tenant)pgvector 0.8.0 iterative scan,halfvec/binary quantization 2025,Amazon Web Services,Inc.or its affiliates.