AI Engineer Summit 2025
Anchoring Enterprise GenAI with Knowledge Graphs
Loading slides…
Read the talk
Anchoring Enterprise GenAI with Knowledge Graphs
A biopharma search application connects document retrieval, organizational incentives and graph relationships to the practical work of moving GenAI into production.
From a talk by Stephen Chin and Jonathan Lowe
Before you start: Basic familiarity with language models and document search is helpful; the article introduces retrieval-augmented generation and graph-based context.
When an executive wants GenAI in production
What happens when an executive sees generative AI handle a personal task and expects an enterprise application to follow just as easily? Stephen Chin opens with the gap between enthusiasm and delivery. He cites Gartner’s forecast that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025. That was a prediction, not a measured outcome. His informal audience poll makes the delivery problem tangible: some attendees acknowledge failing projects, while many others have not yet reached production.
The challenge is to turn that enthusiasm into a technically achievable goal that an organization can understand and support. Chin caricatures an executive whose children use GenAI for school and who therefore expects a business solution in production in two months. The accompanying cartoon makes the same leap from a poem about a dog to business transformation. Jonathan Lowe brings the implementation perspective: building a life-sciences application means navigating leadership, organizational silos and technical constraints together.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
From lab-bench knowledge to factory production
Lowe introduces himself as the skeptical implementer—the person colleagues might think of as Debbie Downer once he explains what launching AI inside a business entails. His concrete use case is technology transfer in biopharma: taking drug development from beakers on a laboratory bench to industrial manufacturing. He illustrates the destination as factories making a million doses a day. Getting there can take years because the people designing factories and equipment must sift through hundreds of thousands of scientific documents, notes and test outcomes. Search is part of transferring the knowledge required to manufacture the product.
The documents are only one part of the knowledge base. Experienced workers also carry tacit knowledge that can disappear when they retire. Lowe cites an unnamed study reporting average manufacturing-worker tenure of about twenty years in 2019 and three years at the time of his account. A matching PMMI account dates the three-year figure to 2023; the underlying population and calculation are not established here. The operational concern is the loss of expertise: incoming workers need access both to what earlier teams wrote down and to what experienced people know but have not documented.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Use the graph to learn what a retrieval unit should be
Lowe describes a corpus of millions of documents represented in a graph, then makes an important distinction: the team did not necessarily load each whole document into the graph. It loaded chunks and preserved their structure—documents, blocks, paragraphs and lines. The slide also shows pages in this hierarchy, alongside medicine and condition nodes. A retrieved fragment could therefore remain connected to the larger document structure it came from.
That structure supported an improvement loop. The team used similarity search, examined which chunks returned what people actually wanted, and refined how chunks were stored and managed. Those observations then informed how subsequent documents were chunked. Chunking was something to learn from retrieval behavior, not merely an ingestion setting chosen once. The graph made different levels of document structure available for that investigation.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
A useful application still has to win adoption
This gives the project a business purpose: reduce the difficulty of recovering knowledge needed for technology transfer, with the potential to get life-saving drugs to people sooner. Chin calls the absence of a real, monetizable use case the largest failure mode in the Gartner study he invokes. That ranking is his characterization; the forecast identifies unclear business value among several abandonment factors. The drug-delivery benefit here is a reason to pursue the application, not a measured clinical outcome.
A good use case does not settle the internal decision. Teams may prefer a different paper, platform or framework to the proposed GraphRAG approach. They may also question the economics. GenAI can be more expensive than conventional computing, particularly when poorly architected, and replacing an existing system requires research, investment and redevelopment. A system that works inadequately can still be difficult to displace if the replacement costs more. Chin frames Lowe’s Pfizer experience around that question: how do you persuade the organization that the improvement warrants the expense?
Lowe approaches the problem as an entrepreneur within a large organization. He describes his current organization as having more than 100,000 people and recalls working at IBM and Deloitte. In the hierarchy on his slide, the person with the idea is a little red figure with a light bulb and a team of data scientists, developers and SREs. Chin suggests going straight to the CEO. Lowe’s instinct is to start with the people who will use the application.
Users apply a practical test:
- Less drudgery: Remove boring work they do not want to do.
- Accurate results: Saving effort is not enough if the answers are wrong.
- Responsive interaction: A button press should not require a coffee break before the result arrives.
Lowe considers satisfying these requirements increasingly achievable. The harder problem is getting the rest of the organization to support the tool.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Translate the tool into the audience’s goals
For someone several levels down in a large company, a meeting with the CEO is unlikely—and an accidental meeting is not necessarily useful. Lowe recalls the scene in Dirty Dancing where Baby, flustered on meeting Johnny, blurts out that she carried a watermelon. He had his own version in a headquarters elevator. The CEO broke the ice, mentioned returning from vacation and asked what Lowe was working on. As the doors opened and the CEO left, Lowe managed: “I'm working with LLMs.” The technology label said little about what the executive wanted to accomplish.
Executive priorities often begin with a broad account of industry leadership and competitive position, shaped in part by consultants. Lowe gives an illustrative purpose blueprint: change a billion lives a year. That is an aspiration, not a result. A useful pitch would connect the search application to that purpose rather than simply announce the use of LLMs. As the purpose moves down the hierarchy, each function translates it into its own goals.
| Audience | Language of value |
|---|---|
| Chief digital officer | Lead the industry in AI |
| Chief scientific officer | Take on the biggest diseases |
| Chief supply officer | Accelerate supply |
| Level-two and level-three leaders | Savings, avoided costs, earlier revenue, balanced head count |
The leaders a project team is more likely to meet need commitments they can use. Slides should connect the proposed capability to numbers and times: what changes, by how much, and when. The same application needs a business explanation at each level of the organization.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Negotiate the scope before it becomes nothing—or everything
Even those business leaders may be behind another organizational boundary. Some companies use client partners to connect digital teams to business units. Because those partners tend to stay within departments, the same search proposal can encounter opposite responses:
- R&D: The department already has five, six or ten search engines. Why build another?
- Supply: Search sounds useful. Why not embed it into every tool in the organization?
One response eliminates the project; the other expands it beyond its original scope. The team has to negotiate both while preserving a deliverable that addresses the original problem.
The next challenge is build versus buy. A vendor may offer a packaged tool while discussing its economics directly with the chief digital officer. After that comes what Lowe calls friendly fire: colleagues may claim AI search as their existing territory, or ask the new team to integrate with everything they have already built. These are different objections—economic justification, ownership and integration—and they need different answers.
Lowe remains optimistic about the opportunity in life sciences and information technology. Concern about failure is part of working near the beginning of a major change. His practical advice is to know each audience and personalize the explanation. The human doing the pitching needs to speak at the right organizational level just as carefully as the application needs to respond appropriately to its users.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Graphs help people understand the data, too
Organizational support does not remove the technical requirement to produce useful answers. Chin returns to hallucinations in enterprise retrieval-augmented generation, or RAG. Newer models and vector retrieval can help supply relevant information, but selecting the right data architecture still matters. Why use a graph at Pfizer?
Lowe’s initial reasons were existing relationship-heavy problems: genealogies of recipes, social networks, hierarchies and time series. As more data was consolidated in the graph, another benefit emerged. Data scientists, engineers, developers and SREs could understand the data landscape more quickly. Lowe reports that consolidating, understanding and cleaning data for a new project fell from three months to three weeks or less. This is his team’s reported experience, without a supplied controlled comparison. The benefit extends beyond traversal or query performance: representing relationships also helped the people building applications make sense of the data.
Chin describes the Pfizer application at this point as an early adopter already in production, although he does not specify rollout timing or scope. He also credits Microsoft with a seminal GraphRAG paper. The matching work, From Local to Global: A Graph RAG Approach to Query-Focused Summarization, extracts entities into a graph and builds community summaries for broad questions across document collections. Its reported improvements concern comprehensiveness and diversity on global sensemaking questions over roughly million-token datasets. That is more specific than a general claim of improved factual accuracy, and its community-summary method should not be assumed to describe Pfizer’s implementation.
Chin positions the choices as a progression in available context:
| Approach | Context supplied | Limitation or intended benefit |
|---|---|---|
| Direct LLM use | Model’s existing knowledge | Missing enterprise-specific information |
| Baseline vector RAG | Similar organizational passages | Answers can remain generic or hallucinate |
| GraphRAG | Retrieved information and explicit relationships | Intended to support more precise, contextual answers |
This is an architectural comparison, not a supplied accuracy benchmark. Life sciences and manufacturing make the distinction consequential because wrong answers can affect business-critical work.
Lowe explains the graph’s contribution through neighborhood retrieval. A relational database can represent connections, but teams may not have made the relevant joins persistent or readily available. In a graph, the modeled relationships are already present. Retrieving one item exposes nearby related information that can also be given to the LLM. The retrieval unit can therefore grow from an isolated matching passage into a connected set of context.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Combine similarity with relationships, then govern the context
The closing architecture maintains both vector and knowledge-graph representations of the data. Its retrieval path is straightforward:
- Retrieve by similarity. Use the vector representation to find information relevant to the user’s request.
- Expand through relationships. Retrieve relationally close nodes from the knowledge graph.
- Assemble the prompt. Pass the retrieved information and additional context to the LLM.
- Return the response. The application presents an answer informed by that combined context.
The diagram places the GenAI application between the user, Neo4j’s graph and vector representations, and an LLM API. Similarity identifies promising information; graph relationships supply additional context around it.
For a small Cypher example of the expansion step, suppose vector retrieval has returned chunk ID paragraph-17. In a teaching graph where that chunk has PART_OF relationships to its enclosing structure, this query preserves the retrieved text and collects its ancestors:
cypher
MATCH (hit:Chunk {id: 'paragraph-17'})
OPTIONAL MATCH (hit)-[:PART_OF*1..3]->(ancestor)
RETURN hit.id AS chunkId,
hit.text AS retrievedText,
collect(DISTINCT {
id: ancestor.id,
text: ancestor.text
}) AS surroundingContext;
The labels, ID and relationship name illustrate the retrieval pattern rather than specify Pfizer’s schema. The query assembles context for a later prompt; it does not generate an answer. In a richer graph, the useful neighborhood could include domain relationships as well as document structure.
The graph can be developed over time or constructed with help from an LLM. Chin then extends the architecture to governance: properties and controls on graph nodes can represent who should have access to particular information. To make that effective, retrieval must enforce those rules before information enters the prompt. A node property is a place to express an access policy, not enforcement by itself.
Explicit relationships also make the retrieved context easier to inspect. People can examine nodes and edges to understand why information is connected—for example, whether a term is related to manufacturing or is merely a general term. That provides an explanation of the context supplied to the model, rather than a complete account of how the model generated every part of its answer.
For technology transfer, these details serve a concrete goal: help people recover the knowledge needed to make medicines and get them to patients sooner. Chin closes on that potential benefit. The graph, retrieval system and organizational work matter because they can support that application—not because adopting GenAI is itself the outcome.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
From the talk
The original Graph RAG paper explains entity graphs and community summaries for answering broad questions across document collections.
Further reading
PMMI discusses workforce retention, knowledge transfer and training technologies in packaging and processing.
Updates since the talk
Current installation and indexing tutorial for experimenting with Microsoft's GraphRAG implementation, including resource-cost warnings.
Full transcript
Loading transcript…