From Retrieval to Action: Building Production Agents on AWS | Milvus Meetup SF

At the Milvus Meetup in San Francisco, Amit Sharma (Senior Manager at AWS Partner Organization) presented “From Retrieval to Action: Building Production Agents on AWS”. In this session, Amit breaks down where AWS and Milvus connect across modern generative AI and agentic workflows, sharing architectural best practices and production lessons learned directly from enterprise customers. He walks through the complete end-to-end agentic pipeline: from upstream embedding generation using Amazon Bedrock and Amazon SageMaker, to vector retrieval engine evaluation criteria, and down to the runtime infrastructure required to power multi-agent systems at scale. Finally, Amit showcases a real-world case study with Rexera, explaining how they automated 99% of complex real estate document workflows using Milvus, LangChain, and LangGraph on AWS. What you'll learn: The end-to-end architecture of Agentic AI pipelines with AWS and Milvus Upstream embedding generation: Amazon Bedrock (managed) vs. Amazon SageMaker (custom models) Key upstream best practices: splitting query and corpus ingestion paths, vector versioning, and dimension optimization Deploying embeddings safely with blue/green cutovers and keeping vectors near the data source How to evaluate retrieval engines for agentic AI: tail latency, cited deduplicated hits, and hybrid semantic/keyword search Downstream agentic architecture: understanding in-context conversation memory vs. long-term vector memory AWS Agentic Harness & Runtime: session isolation, multi-agent concurrency, and security guardrails Production Case Study: How Rexera automated 99% of document cross-referencing workflows with Milvus on AWS Chapters: 0:00 Introduction & The AWS Perspective 1:08 End-to-End Pipeline: From S3 Storage to Agentic AI 2:23 Upstream Embeddings: Amazon Bedrock vs. Amazon SageMaker 4:34 Best Practice 1: Split Query & Corpus Ingestion Paths 5:36 Best Practice 2: Version Every Vector with Model IDs 6:11 Best Practice 3: Dimension Tradeoffs & Blue-Green Embedding Deployments 7:11 Best Practice 4: Keep Vectors Next to the Source 7:50 Evaluating Retrieval Engines for Agentic AI 8:45 Low Tail Latency at High Concurrency & Longest-Pole Bottlenecks 9:15 Compact Cited Hits, Hybrid Filtering & Replayability 10:50 Downstream Architecture: Agent Loops & Memory Types 12:30 The Agentic Harness: Frameworks, SDKs, and Context 14:04 Agent Runtime: Session Isolation, Multi-Agent Concurrency & Guardrails 15:45 Customer Case Study: Rexera Automating 99% of Workflows with Milvus 17:02 Conclusion & Wrap-Up This talk was recorded at the Milvus Meetup in San Francisco (525 Market St). The event brought together search, ML infrastructure, and data platform engineers to explore the algorithms, infrastructure, and production patterns behind modern AI-powered retrieval systems. Watch the full event recording: Milvus Meetup San Francisco October 2026 (Full Event): https://youtu.be/itISrq5J-Ik Watch other talks from this meetup: Milvus 3.0: A More Powerful Retrieval Engine: https://youtu.be/j9qjA73Jj-I Large Scale Entity Resolution with Milvus at 6sense: https://youtu.be/L7DhSBXXomo Find us on: Zilliz Website: https://zilliz.com Milvus Website: https://milvus.io Zilliz LinkedIn: https://www.linkedin.com/company/zilliz/ Milvus LinkedIn: https://linkedin.com/company/67143129 Milvus X: https://x.com/milvusio Milvus GitHub: https://github.com/milvus-io/milvus Try Zilliz Cloud: https://cloud.zilliz.com/signup