Agent Search – AI-Powered Semantic Enterprise Search on AWS

Executive Summary 

Enterprises often face challenges in retrieving accurate and unified information across disparate data systems. Agent Search addresses this by combining AWS-native services such as Amazon EKS, Bedrock, Neptune, OpenSearch, and RDS to deliver semantic search capabilities over both structured and unstructured data. Built on GraphRAG architecture, the platform leverages LLM-powered reasoning with real-time context retrieval from graph and vector stores. This case study outlines how Agent Search helped customers unify access to knowledge silos, enhancing productivity, compliance, and decision-making across departments. 

Customer Challenge 

Customer Information 

• Customer: Confidential (Representative of BFSI, Healthcare, and E-commerce sectors) 

• Industry: Multi-industry (BFSI, Healthcare, E-commerce) 

• Location: Global operations 

• Company Size: 1,000+ employees 

Business Challenges 

• Data scattered across relational databases, search indexes, object storage, and knowledge graphs created a fragmented experience for business analysts and compliance teams. 

• Existing tools offered only keyword-based search with no semantic understanding, limiting their effectiveness for legal, policy, and customer insights. 

• Stakeholders lacked traceable, accurate answers to business-critical questions. 

• Regulatory requirements demanded precise documentation access with traceability. 

• There was increasing pressure to enable real-time policy discovery, especially in fast-changing compliance environments. 

Technical Challenges 

• Integrating and indexing data from multiple sources: S3, RDS, Neptune, and OpenSearch. 

• Ensuring secure IAM-based access while maintaining auditability and compliance. 

• Enabling performant semantic search at scale with scalable, containerized infrastructure. 

• Operationalizing GraphRAG on Kubernetes while managing LLM interactions through Bedrock. 

• Aligning with AWS security and VPC policies for enterprise compliance.

Partner Solution 

Solution Overview 

Agent Search delivers an intelligent, LLM-powered, real-time semantic search platform built on Kubernetes and AWS-native services. The platform enables cross-silo search by integrating graph databases (Neptune), vector stores (OpenSearch), and relational data (RDS), enhanced through Bedrock LLMs. A modular indexing framework built on Kubernetes jobs handles ingestion, while GraphRAG enables dynamic retrieval-augmented generation (RAG). IAM-secured APIs allow seamless integration across the enterprise. 

AWS Services Used 

• Amazon EKS: Hosts the GraphRAG API server and indexing workers. 

• Amazon Bedrock: Provides LLM reasoning capabilities. 

• Amazon Neptune: Stores graph data and metadata context. 

• Amazon OpenSearch: Used as a vector DB for semantic similarity search. 

• Amazon RDS: Stores structured data such as policy tables. 

• Amazon S3: Houses documents and unstructured files. 

• IAM: Controls secure, scoped access to services. 

• CloudWatch: Enables observability and alerting. 

• Amazon ECR: Container image repository for deployment. 

Architecture Diagram 

Implementation Details 

• Deployment: Helm-based deployment on Amazon EKS using separate control and data planes. 

• Methodology: Agile delivery with continuous integration and testing pipelines. 

• Security: IAM roles scoped per service with TLS in transit and SSE encryption at rest. 

• Testing: Performed synthetic and real-data search scenarios with trace logging. 

• Timeline: Initial prototype in 2 weeks; production deployment in 6 weeks. 

• Integration: API Gateway enabled access to web UIs and Slack bots. 

• Observability: CloudWatch and custom dashboards ensure transparency in query resolution. 

Innovation and Best Practices 

• Used GraphRAG with AWS-native vector and graph stores for real-time contextual augmentation. 

• Employed Kubernetes autoscaling for indexing scalability and latency reduction. 

• Followed AWS Well-Architected Framework: security, reliability, performance efficiency. 

• Integrated CI/CD pipelines to manage iterative improvement cycles and LLM prompt tuning. 

• Introduced reusable Helm charts and IAM templates for faster customer rollout. 

Results and Benefits 

Business Outcomes and Success Metrics 

• Reduction in Time to Insight: 65% faster discovery of policy and compliance documents. 

• Increased Query Accuracy: 80% improvement in answer relevancy compared to legacy search. 

• Improved Compliance: Enabled traceable document access for audits and regulators. 

• Team Productivity: Reduced analyst research time by ~40% via semantic automation. 

• Cost Efficiency: Eliminated siloed tools, reducing TCO by ~30%. 

Technical Benefits 

• Elastic Scalability: EKS-based worker pods scale with query load. 

• Performance: Optimized through async indexing and Bedrock’s LLM streaming. 

• Availability: High availability via multi-zone Kubernetes deployment. 

• Security: Full audit trail with scoped IAM, VPC isolation, and encrypted data layers. 

• Modular Extensibility: New sources and use cases can be added with minimal rework. 

Lessons Learned 

Challenges Overcome 

• Initial ingestion of large legacy datasets required batch migration jobs and tuning of indexing pods. 

• Managing Bedrock API limits requires implementing retry and fallback strategies. 

• Ensuring consistent context mapping across graph/vector stores requires schema unification.

Best Practices Identified 

• Pre-indexing data during off-peak hours reduces latency under heavy query load. 

• Establish robust observability (CloudWatch + custom traces) for accelerated debugging. 

• Modularizing the agent pipeline helps isolate and tune RAG components independently.

Future Plans 

• Expand deployment to multiple regions and introduce multi-tenant features. 

• Add support for user feedback loops to tune model outputs. 

• Integrate with Microsoft Teams and Confluence for inline search. 

• Explore AWS Marketplace launch and open ecosystem partner integrations.