Building Production Multi-Agent RAG Systems on Serverless AWS
summary
Claims deployment of a production multi-agent RAG platform on serverless AWS with a five-agent financial analysis system called Prime and concrete lessons about what actually worked and failed. Uses SQS + Lambda orchestration, Bedrock multi-region routing for LLMs, SageMaker embeddings, S3 vector store vs OpenSearch tradeoffs and achieves 90% storage savings on vectors in this workload. Covers cost/latency numbers, cross-region LLM routing, observability with dead-letter queues & Langfuse and practical fixes for Bedrock rate limits and Lambda logging timeouts.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 3:02Single-agent limitations
- 5:51Prime multi-agent system
- 7:18Architecture and SQS orchestration
- 10:50SQS resilience
- 12:14Strategic RAG in agents
- 13:37Bedrock cross-region inference
- 15:47Cost-effective S3 storage
- 17:52Observability importance
- 21:26Key takeaways
related talks
29mBuild and scale reliable multi-agent systems on AWS
41mServerless LLM with AWS Lambda and LangChain: A Revolution in Application Development
15mServerless & Agentic AI: Better Together
28mBuilding Agentic AI applications on AWS
23mBuilding Agentic AI Systems with AWS Serverless
40mBuilding an AWS solutions architect agent with Amazon Bedrock
01h 02mServerless in production, an experience report
47mServerless in production, an experience report