Knowledge Graphs, Neuro-Symbolic Practice, and Grounding LLM Context

For RMAIIG's August AI/ML Engineering session, let's talk about knowledge graphs and where they meet modern AI practice.

Uche Ogbuji opens with an introduction to the topic, aimed at people who know LLMs well and graph data systems less well. Uche has been working with KGs since discovering RDF in 1999, through the early schema.org effort, and is revisiting with new urgency and enthusiasm, in how to give language models context structured enough to reason over.

LLMs are fluent and tolerant of ambiguity but hold no commitment to ground truth, Symbolic representations are explicit, compositional, and auditable but can't handle the messy edges. Neuro-symbolic practice is the attempt to get both halves of the digital brain working together. Knowledge graphs are the most practical symbolic substrate available. Microsoft's GraphRAG made the popular case that relational structure beats flat chunk similarity for multi-hop questions, but an even deeper win is provenance: when a graph carries where each assertion came from, you can audit an answer rather than just admire it.

Uche's talk will also touch on Onya, his open-source KG library, which takes a Markdown-native approach to serialization so graphs stay readable and diffable by humans while also remaining very digestible to LLMs. He'll show some of Onya's visualization work. He closes with a live demo of the @vv tool from a sandbox tenant. @vv is a knowledge-graph-backed agent platform, and helps demonstrate what this all looks like when it's carrying real weight.

Additional speakers to be announced.

Hope to see you there, and as ever, I'll bring the coffee!