Success Story

• The client is an independent statutory corporation responsible for operating and maintaining a provincial land title and survey system. It provides secure, legislated land registration, survey access, and parcel mapping services to professionals, government agencies, and the public.
• Its web-based Help Centre serves as the primary self-service channel for registration guidance, account setup, and search services.

The existing chatbot was built on a rigid intent-based NLP architecture with limited coverage and high manual maintenance overhead.

The client required a next-generation Generative AI assistant to —

• Handle open-ended, natural language queries across full Help Centre content
• Expand knowledge coverage beyond limited predefined intents​
• Reduce manual maintenance through managed RAG architecture​
• Improve response speed and contextual accuracy while maintaining regional compliance​

A production-grade Generative AI chatbot was deployed using Amazon Bedrock and a fully managed retrieval pipeline.

• Managed Knowledge Base – Amazon Bedrock Knowledge Base integrated with Amazon OpenSearch Serverless replaced the evaluated self-hosted vector database. Amazon Titan Text Embeddings V2 enables semantic retrieval.
• Foundation Model Selection – GPT-4o-mini, Claude 3 Haiku, and Claude 3 Sonnet were evaluated across latency, accuracy, cost, and Canada region availability. Claude 3 Haiku was selected for low latency, high context window (up to 200K tokens), and regional compliance.
• Relevance Optimisation – Cohere Rerank 3.5 refines retrieved results before response generation.
• Backend Architecture – Containerised AI-API deployed on Amazon EKS with health probes and ReplicaSets for self-healing and high availability. Secure Bedrock connectivity via AWS PrivateLink.
• Frontend & Storage – UI hosted on Amazon S3 with CloudFront and AWS WAF. Chat history stored in Amazon DynamoDB (24-hour TTL). Amazon RDS (PostgreSQL) supports trace storage and replication for DR.
• Guardrails & Observability – NeMo Guardrails enforces topic-level allow/deny controls. LangFuse captures LLM traces. GlitchTip and Imperva provide monitoring and application security.

• Full replacement of intent-based chatbot with conversational Generative AI assistant
• Improved response speed via low-latency Bedrock inference
• Enhanced contextual accuracy through managed RAG and reranking
• Increased self-service engagement across Help Centre users
• Reduced support workload through automated query resolution