A MENA commerce platform replaced slow, imprecise search with a private, VPC-only GenAI experience on AWS—combining Amazon Bedrock with hybrid OpenSearch and RAG on S3. The result: sub-second search and on-topic answers that help shoppers find the right product on the first try.
Regional multi-store retailer operating in Bahrain and the GCC, with a Bahrain-hosted SaaS commerce front end.
Shoppers faced slow product search (>2s P95) and off-target results; support teams handled repetitive policy questions (shipping/returns). The customer needed a secure, low-latency GenAI experience integrated with their existing Bahrain VPC, without public egress.
Solution
GenAI orchestration with Amazon Bedrock
Claude for conversational reasoning; Cohere Embeddings for retrieval; Titan for on-demand image generation. Accessed via VPC Endpoints (no public internet)
Hybrid search on Amazon OpenSearch Service
BM25 + vector k-NN (HNSW) with rank fusion; blue/green index releases via aliases
RAG on Amazon S3 (LanceDB-backed)
Policy/FAQ content chunked, embedded, and upserted idempotently by content hash
Serverless integration
AWS Lambda for routing, retrieval, and generation; DynamoDB (chat history/config); ElastiCache Redis (low-latency caching)
Networking
Bahrain VPC ↔ us-east-1 VPC peering; interface/gateway endpoints for S3, Bedrock, DynamoDB, Lambda, OpenSearch
Results
Faster answers
Time-to-first-result dropped to ~0.5s median (P95 <1s)
Higher relevance
Top-1 intent match rose to ~84%, Success@3 ~95%; near-zero zero-result searches
Lower support load
Policy Q&A deflected to self-service with grounded responses
Operational safety
Zero-downtime search releases via OpenSearch aliases; weekly prompt/index evaluations
AWS Services Used
Amazon Bedrock (Claude, Cohere Embeddings, Titan), Amazon OpenSearch Service, AWS Lambda, Amazon DynamoDB, Amazon ElastiCache for Redis, Amazon S3, Amazon AppSync/API Gateway, Amazon SageMaker (evaluation), VPC Peering & VPC Endpoints, Amazon CloudWatch.
Why AWS
AWS provided managed foundation models and secure private connectivity, letting the team deliver grounded GenAI quickly while keeping costs and operations low. Bedrock’s VPC integration and OpenSearch’s hybrid retrieval reduced complexity versus self-hosting or third-party APIs.