Novo Nordisk Accelerates Process Documentation Answers With Amazon Bedrock AgentCore and GraphRAG
Novo Nordisk's manufacturing team has deployed an AI process-intelligence assistant on Amazon Bedrock AgentCore, leveraging a GraphRAG architecture to deliver grounded, auditable answers from a large set of regulated documents.
What changed?
Novo Nordisk has introduced a process-intelligence assistant for regulated manufacturing processes, built on Amazon Bedrock AgentCore. The assistant uses a GraphRAG architecture that combines a knowledge graph in Amazon Neptune with hybrid semantic and exact match search via Amazon OpenSearch Serverless. AI agents powered by Anthropic Claude, orchestrated with AgentCore, manage end-to-end retrieval, clarification, and answer generation. The system provides grounded, cited responses to complex process queries and explicitly identifies when information is unavailable. Ontology-guided extraction with a fixed, expert-curated schema underpins auditability and regulatory compliance.

Why does it matter to an everyday developer?
This deployment highlights practical advances for developers working with large, regulated document sets or other enterprise knowledge bases. The assistant demonstrates how hybrid architectures—marrying knowledge graphs with semantic and lexical search—overcome the limitations of plain vector-based RAG, particularly for retrieving content by exact identifiers and resolving multi-step relationships. It also shows the growing best practice of using predefined ontologies and schema for extraction, which improves trust, traceability, and auditability, all crucial in regulated settings. Guided search that clarifies ambiguities before retrieval minimizes both irrelevant and potentially inaccurate responses, raising answer reliability.
What can the developer do now?
Developers working in regulated industries or with highly structured document collections can study the GraphRAG approach to address similar challenges. Teams can architect AI-powered assistants using a combination of a relational or graph store for structure, a search index for document recall, and an orchestration layer for reasoning and guided user interaction. Implement ontology-guided extraction with a defined, expert-maintained schema for domain entities and relationships to support compliance and auditability. The AWS reference solution, including AgentCore and open components like the Strands Agents SDK, offers a template to adapt. Testing guided retrieval workflows that clarify intent before fetching documents can improve both relevance and user trust in outputs.
