Novo Nordisk Deploys a GraphRAG Assistant for Regulated Manufacturing on Amazon Bedrock AgentCore
Summary:Novo Nordisk's Manufacturing Technology team has deployed a process-intelligence assistant on AWS for a vault containing roughly four million controlled documents. Amazon Bedrock AgentCore runs the query agent; Amazon OpenSearch Serverless combines BM25 with Titan dense embeddings, while Amazon Neptune stores the process hierarchy, manufacturing entities, document lifecycle state, and cross-document relationships. Amazon Bedrock Rerank then refines the fused candidates before the agent answers w
Novo Nordisk's Manufacturing Technology team has deployed a process-intelligence assistant on AWS for standard operating procedures, work instructions, forms, and training materials used in regulated manufacturing. The central design choice is to keep document retrieval, scope resolution, version governance, and citation verification inside an auditable agent workflow instead of asking a model to answer directly from its own knowledge.
Core Architecture: More Than a Graph Beside a Vector Store
The system separates relational structure from document content and connects the two stores at query time:
- Amazon Neptune stores a two-layer taxonomy. The upper layer is an approximately three-level process tree, with each controlled document attached to its most specific process node. The lower layer models sites, product generations, equipment, and process steps as entities connected through a small set of unified relationships.
- Amazon OpenSearch Serverless handles content recall. Amazon Titan Text Embeddings provide dense semantic retrieval, while BM25 captures exact tokens such as document numbers, equipment codes, and multilingual terminology.
- A shared document number is the deterministic bridge between the graph and the full-text index. An OpenSearch hit can enter the corresponding Neptune node directly, without a second fuzzy entity-matching pass.
- Amazon Bedrock Rerank applies a cross-encoder after candidate retrieval and graph expansion, scoring the question and each candidate passage jointly.
Hybrid recall uses Reciprocal Rank Fusion. A typical formulation is RRF(d) = Σr 1/(k + rank_r(d)). RRF combines ranks rather than trying to normalize vector similarity and BM25 scores into a common scale. The smoothing parameter, candidate limits, and retrieval weights still require tuning against domain questions.
Query Path: From Scope Clarification to Cited Answers
- The query agent determines whether the question has a unique applicable context. A request such as “which samples are taken on the granulator?” may map to different sites, product generations, equipment, or batch sizes, making immediate retrieval prone to confident scope errors.
- Guided Mode resolves universal attributes such as site and product generation first, then derives deeper options such as equipment and batch from Neptune. Every clarification option comes from an existing graph node rather than model-generated speculation.
- OpenSearch performs BM25 and dense retrieval, after which graph expansion adds neighboring documents, references, process hierarchy, and change-impact relationships.
- Bedrock Rerank refines the candidates before Claude models running on Amazon Bedrock compose the response.
- By default, the assistant uses approved, effective, and applicable versions only. When evidence is missing or ambiguous, it states that limitation and flags the case for subject-matter expert, quality, or regulatory review.
Amazon Bedrock AgentCore provides the managed agent runtime, including session isolation, tool orchestration, IAM integration, and observability. The agents are implemented with the Strands Agents SDK. For ingestion, Amazon Bedrock Data Automation performs multimodal parsing, AWS Step Functions and AWS Lambda orchestrate the workflow, and AWS CDK manages the infrastructure as code. The case study does not identify the exact Claude model version, preventing a reliable assessment of the quality-cost tradeoff.
Grounding for Regulated Manufacturing
The architecture promotes metadata that conventional RAG pipelines often treat as an afterthought into first-class constraints:
- Approval status, effective state, lifecycle status, and applicability to sites, products, and equipment are represented explicitly in the graph.
- A separate
indexedstate distinguishes “the document and its metadata are known” from “the document's full text has been indexed.” An empty search result is therefore not mistaken for proof that the knowledge does not exist. - Answers must preserve requirement strength, distinguishing
must,should, andmayso that recommendations are not rewritten as mandatory instructions. - The automated evaluation harness checks source relevance, document status, version applicability, scope compliance, citation accuracy, and preservation of requirement strength. These are more meaningful acceptance criteria for this use case than answer-string similarity alone.
Engineering Trade-offs: Precision, Latency, and Maintenance
| Design choice | Primary benefit | Engineering cost or risk | |---|---|---| | Separate Neptune and OpenSearch stores | Each workload uses an appropriate query model, and relational structure is not forced into a vector index | Dual-write consistency, event propagation delays, node deletion, and index rebuilds must be handled; query-time bridging adds another service interaction | | Fixed, expert-curated ontology | Stable entity types and auditable extraction suit GxP document governance | New domains and entities require SME definition, human promotion, and prompt changes; the model cannot invent schema freely | | Clarify before retrieval | Narrows the applicable scope and reduces confidently incorrect answers | Simple questions may require an extra interaction; the UI must expose candidate scopes and why they were requested | | Chained BM25, vectors, graph expansion, and reranking | Covers exact identifiers, semantic recall, and multi-hop relationships | Multiple stages increase latency and cost; candidate count, graph depth, and reranking batch size need hard limits | | Managed AgentCore runtime | Reduces duplicated work for sessions, tools, and observability | Cloud-platform dependence remains; IAM granularity, audit export, regional compliance, and model data residency require separate validation |
Graph traversal in particular needs a strict budget. Unbounded multi-hop expansion can turn a direct question into a large subgraph query, increasing latency while introducing weakly related documents. Production systems should make traversal depth, fan-out, candidate passage count, and reranking batch size configurable, while recording how each piece of evidence entered the final context.
Performance Evidence and Evaluation Boundaries
Novo Nordisk and AWS report that search, document-update, and context-adaptation tasks that previously took hours or days can now be completed in minutes. The system can also trace reference chains and identify links to obsolete or superseded material. These results indicate that the architecture reached a real operational workflow, but the article provides no accuracy rate, refusal rate, citation-correctness score, p50 or p95 latency, token consumption, AWS cost, indexed-corpus coverage, or matched comparison against traditional search and vector-only RAG.
The automated evaluation framework lists appropriate dimensions, but it does not disclose test-set size, question mix, human annotation procedure, or score distributions. If the model, taxonomy, and acceptance questions are repeatedly tuned on the same cases without separate tuning and holdout sets, overfitting to fixed scenarios becomes possible. The deployment is therefore a valuable production architecture reference, not a portable performance benchmark.
Deployment Guidance
- Govern document keys and lifecycle first. Document ID, version, approval state, effective state, applicability, and full-text indexing status should be structured during ingestion rather than inferred by the model.
- Use deterministic join keys. Sharing stable IDs between search passages and graph nodes avoids incorrect links caused by entity renaming and fuzzy matching.
- Choose an entry path based on question type. For semantic questions, retrieve vectors first and then expand the graph; for relationship or impact analysis, begin from graph entities.
- Adopt scope-first behavior for high-risk questions. Resolve site, product generation, and equipment before retrieval, while retaining a no-clarification fast path for low-risk, routine requests.
- Separate unindexed from unanswered. The interface and agent should expose evidence-coverage state instead of returning an undifferentiated empty result.
- Evaluate requirement-level correctness. In addition to answer correctness, track source selection, status filtering, scope compliance, citation support, requirement-strength preservation, and confidently incorrect responses as separate metrics.
- Keep a human escalation path. Missing evidence, conflicting versions, or regulation-sensitive clauses should route to subject-matter experts or quality reviewers rather than forcing the model to produce a conclusion.
The Novo Nordisk and AWS collaboration demonstrates that production GraphRAG in regulated manufacturing is not simply about adding a graph to retrieval. Its value comes from integrating scope resolution, lifecycle governance, explicit indexing state, and auditable refusal into the agent workflow. The four-million-document scale is a strong motivation for the architecture, but production teams still need to validate performance and cost against their own question sets.
— END —Source: AWS Machine Learning Blog (2026-11-16)
Tags: #Amazon Bedrock AgentCore #GraphRAG #Amazon Neptune #Hybrid Retrieval #Pharmaceutical Manufacturing
Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.
Community Comments