LLMs also prefer stories to graphs and databases

Before starting Lightfield, we tried many different approaches to context engineering and data models. Unsurprisingly, we found that most models (like humans) responded best to stories and narratives rather than tables and graphs.

The human proof is simple. Imagine onboarding someone at your company. Which of these would work best?

• Dump a spreadsheet on them and ask them to make sense of it
• Build an arbitrary knowledge graph connecting concepts semantically and ask them to analyze it
• Tell them a story that pulls together the artifacts and data into a narrative

It should make intuitive sense. It's the reason people have relied on biblical stories to communicate values instead of datasets.

When we used LLMs with structured graphs, the models treated the relationships too rigidly. They'd hold the edges too firmly and lacked understanding ofwhyconcepts were connected or what the overarching theme was across the data. They were missing the forest for the trees.

I don't think most folks have run into this problem with knowledge graphs for the average LLM application. You run into it when you have to understand something as complex as human relationships. Modeling human relationships requires changing the weights of details to fit the dominant narrative. Maybe the CISO you were trying to sell to said he had a fixed budget and no time for you, but then you explained you could solve a bigger problem, and his view on budget and urgency changed. Graphs really got in the way of modeling the malleability of humans.

In contrast, if you write up all the details about a person and your relationship with them as a story, the best models (e.g., Opus) can re-weight the importance of details dynamically. This is true of humans too. Give them the chronology and the narrative and they'll find the global optimum instead of getting stuck in the details.

So last year, we dumped the classical graph and started building our data model of people and companies around stories. We write live-updating chronological stories about people and your relationship with them, feeding them to models alongside all the structured and unstructured data you'd expect. The models do much, much better with this kind of input.

Here's an illustrative example: if you do enterprise sales, you often need to navigate 50+ relationships and come up with a plan of persuasion to win over everyone in the right order with the right incentives. A rigid context graph gets in the way of understanding all of these humans and how their priorities and minds change. In the video below, we're able to nail this high-complexity task: a stakeholder map, engagement analysis, risk analysis, next steps, and deal outlook. We also fetch all the data through code execution to make sure it's not lossy.

There's a lot more power to unlock by making context engineering look more like persuasive storytelling.