The Benchmark

Novo Nordisk's GraphRAG result deserves a calibrated read

A pharma manufacturer rebuilt process intelligence with a knowledge graph. The architecture tells you more than the press release does.

AIBy The Beacon desk, Lighthouse Strategy

Data pulse

1

Pharma manufacturer running GraphRAG on process docs

Source: AWS Machine Learning Blog

Novo Nordisk just published a case study on its manufacturing process intelligence system, and the architecture is worth reading slowly. The company's Tablet area team built an assistant on Amazon Bedrock AgentCore using a GraphRAG design. That is retrieval-augmented generation layered onto a knowledge graph rather than a flat vector index. If you have watched AI deployments stumble on documents with dense cross-references, the distinction matters.

What GraphRAG actually does differently

Standard RAG works by converting documents into vector embeddings and retrieving the chunks closest to a query. It is probably fine for FAQs. It is not fine for manufacturing process documentation, where a step in one procedure inherits a constraint from a document written three years earlier by a different team. A flat vector index has no reliable way to surface that relationship. A knowledge graph does. GraphRAG uses the graph structure to traverse relationships explicitly before passing context to the model. That reduces the probability that the model hallucinates a process step because it retrieved the wrong chunk.

Novo Nordisk's document corpus is roughly the right shape for this approach. Pharmaceutical manufacturing operates under strict procedural dependencies. Equipment qualifications reference materials specifications. Batch records reference equipment qualifications. The graph captures that web. Your document corpus may not have the same density of explicit relationships, and that is an honest constraint worth naming before you plan a similar build.

The parts the press release skips

Every case study from a cloud vendor has a selection problem. You are reading about the deployment that worked. The ones that stalled in a pilot are not published. That is not a reason to dismiss the Novo Nordisk result. It is a reason to pressure-test whether your conditions match theirs before you commit a team to a six-month build.

Three conditions probably explain why this worked. First, the documents have a controlled vocabulary. Pharmaceutical manufacturing uses defined terminology enforced by regulatory bodies. Your internal product documentation probably does not have the same discipline. Second, the team built the knowledge graph as a deliberate upstream step, not as an afterthought. Graph construction requires a schema. That schema requires subject matter experts to define entity types and relationship types before any model is involved. That work is unglamorous and expensive. Third, the deployment runs on a managed layer, Amazon Bedrock AgentCore, which reduces infrastructure latency and simplifies the operational surface. Vendor lock-in is the tradeoff. If AWS changes pricing or deprecates a feature, the migration cost is real.

Where commerce teams could borrow this logic

You are probably not running pharmaceutical process documentation. But you likely have document sets with similar structural properties. Product specification sheets that reference compliance certifications. Supplier contracts that reference materials standards. Catalog data where a product's eligibility for a promotion depends on a category rule written in a different system. Any corpus where relationships between documents carry operational meaning is a reasonable candidate for a graph-augmented retrieval layer.

The inference worth drawing is narrow and calibrated. GraphRAG is not a universal upgrade to your retrieval stack. It is likely the right tool when two conditions hold simultaneously. Your documents contain many-to-many relationships that matter for the answer. And the cost of a hallucinated answer is high enough to justify the extra schema work. For a commerce team, that probably means contract and compliance documents before it means marketing copy or product descriptions.

Three questions to pressure-test your eval

Before you route this case study to your engineering team, run it through these questions. First, can you name three relationship types that are both explicit in your documents and load-bearing for a business decision? If you cannot name them, your corpus may not warrant the graph layer. Second, do you have subject matter experts who can define a schema before the first model call, or will you try to infer the schema from the documents themselves? The second path usually produces a brittle graph. Third, what is the actual cost of a wrong answer in your target use case? If a hallucinated process step in your context means a delayed order rather than a regulatory violation, the lighter retrieval architecture is probably sufficient and cheaper to maintain.

One uncertainty I cannot resolve from the published case study: Novo Nordisk has not released evaluation metrics on retrieval accuracy before and after the graph layer. The claim that GraphRAG outperforms flat vector retrieval on this corpus is plausible and consistent with the research literature, but the specific numbers are not visible here. What would change my view is a published eval comparing retrieval precision across both architectures on the same document set. Until that exists, the result is a strong signal, not a proof.

From the publisher

Reading is free. Acting on it is where we come in.

When a story is about your category, a Discovery call turns it into a plan.