# Transforming Insurance Agent Support with RAG and Generative AI at InsuranceDekho ## Company Background and Challenge InsuranceDekho is a leading InsurTech service in India that works with over 49 insurance companies through a network of 150,000 point of sale person (POSP) agents. The company faced a significant operational challenge where insurance advisors, especially newcomers, frequently needed to consult subject matter experts (SMEs) for policy-specific queries. This dependency created delays in customer response times and increased operational costs. ## Technical Solution Architecture ### Core Technology Stack - **Foundation Model Platform**: Amazon Bedrock - **Large Language Model**: Anthropic's Claude Haiku - **Vector Database**: Amazon OpenSearch Service - **Caching Layer**: Redis on Amazon ElastiCache - **Third-party embedding model** for vector transformations ### Implementation Workflows The solution comprises two main workflows: ### Data Ingestion Workflow - Policy documents are processed through embedding models - Vector representations are stored in OpenSearch Service - Continuous updates maintain knowledge base currency ### Query Response Workflow - Chatbot interface for insurance advisor interactions - Redis-based caching system for frequent queries - Intent classification using Claude Haiku - RAG-based context retrieval and response generation - Dynamic prompting based on query type ## Key Technical Decisions and Rationale ### Amazon Bedrock Selection - **Serverless Architecture**: Eliminated need for infrastructure management - **Model Flexibility**: Single API access to multiple foundation models - **Easy Model Switching**: Demonstrated by smooth transition from Claude Instant to Claude Haiku - **Security**: AWS PrivateLink integration for private model access - **Performance Options**: Both on-demand and provisioned throughput available ### Performance Optimization Strategies - **Caching Layer Implementation** - **Intent Classification** ### Vector Database Implementation - OpenSearch Service chosen for: ## Production Deployment Considerations ### Security and Compliance - Private model access through AWS PrivateLink - Secure data transfer for inference - Compliance adherence in data handling ### Performance Monitoring - Response time tracking - Cache hit rates - Query classification accuracy - Vector search performance ### Scalability Design - Serverless architecture for variable load handling - Elastic cache scaling - Vector database performance optimization ## Results and Impact ### Performance Improvements - 80% reduction in query response times - Significant decrease in SME dependencies - Enhanced agent confidence and autonomy ### Business Benefits - Improved customer service efficiency - Reduced operational costs - Better scalability of support operations - Enhanced accuracy in policy information delivery ## Technical Learnings and Best Practices ### RAG Implementation - Importance of quality embeddings - Context window optimization - Prompt engineering for accurate responses ### System Integration - Seamless integration between multiple AWS services - Effective caching strategies - Query optimization techniques ### Model Selection and Evaluation - Benchmarking process for LLM selection - Performance metrics consideration - Cost-effectiveness analysis ## Future Considerations ### Potential Enhancements - Continuous model updates and evaluations - Enhanced caching strategies - Advanced intent classification - Expanded knowledge base coverage ### Scalability Planning - Load balancing improvements - Cache optimization - Vector database performance tuning ## Technical Architecture Details ### Data Flow - Query input through chatbot interface - Cache check with semantic search - Intent classification processing - Context retrieval from vector database - Response generation and delivery ### System Components - Frontend chatbot interface - Caching layer with Redis - Vector database with OpenSearch - LLM integration through Amazon Bedrock - Intent classification system ### Monitoring and Maintenance - Performance metrics tracking - Cache hit rate monitoring - Response accuracy evaluation - System health checks - Knowledge base updates