Report

The Connected Intelligence Experiment

What 300+ AI builders learned about combining knowledge graphs and LLMs

Give hundreds of teams the same tools, a set of enterprise problems and no prescribed architecture. What do they build? In this experiment, more than 300 builders did exactly that on TigerGraph. Working independently, the strongest teams kept arriving at the same design choices for connecting knowledge graphs to large language models. This report lays out those principles and the reasoning behind them.

Download the free report

Design principles the strongest teams shared

Model the relationships, not just the documents

Top builds defined the entities and links in their domain before writing prompts.

Let the graph decide what the model sees

Graph queries selected a small, connected set of facts, keeping prompts short and focused.

Use vectors and graph together

Semantic search found the starting points and graph traversal filled in the context around them.

Keep the reasoning visible

Teams that exposed the path behind each answer found errors faster and earned trust sooner.

What the report covers

How the experiment was set up and how builds were scored

The design principles that kept appearing in the strongest builds

Common mistakes that held weaker builds back

Reference patterns for connecting a knowledge graph to an LLM

How to apply the principles to your own AI roadmap

Who should read it

Architects and AI leads designing LLM applications on enterprise data who want patterns proven by many independent teams rather than a single vendor reference design.

Knowledge graphs and LLMs, answered

How do knowledge graphs improve LLM accuracy?

A knowledge graph gives an LLM structured facts and the relationships between them. Instead of guessing from loosely related text, the model works from a connected view of entities, which improves accuracy on complex questions, reduces hallucination and makes it possible to trace how an answer was formed.

How do you use a knowledge graph with an LLM?

Queries run against the knowledge graph, often alongside vector search, to assemble the entities and relationships relevant to a question. That connected context is passed to the LLM in the prompt, and the model generates an answer grounded in those specific facts rather than in its training data alone.

Why run an open experiment instead of a benchmark?

A fixed benchmark tests one architecture. Leaving the design open shows which choices teams make on their own when solving real problems. When hundreds of independent builds converge on the same patterns, that is strong evidence those patterns work.

Build on what already works

Get the full set of design principles from 300+ builds. To try them on your data, book a TigerGraph demo or start free in Savanna.

Download the free report Contact us