Michael Iantosca (Avalara): Graph-Driven RAG AI Powered by DITA

This talk was presented at the AI The Docs 2025 online conference. We are thrilled to share the recording and the summary with you. 

Visit the talk summary page to see all of the presentations from the conference.

presentation

Michael Iantosca (Senior Director of Content Platforms and Knowledge Engineering at Avalara)

Summary

How can AI achieve true deterministic reasoning, reliability, and provide robust context for content? What specific advantages do knowledge graphs offer for building trustworthy and advanced AI applications?

In his presentation, Michael Iantosca (Senior Director of Content Platforms at Avalara Inc.) explains that today’s AI models (LLMs and vector-based RAG systems) are fundamentally unreliable because they lack deterministic reasoning, discard content context, and rely only on predictive outputs. Knowledge graphs were presented as the missing foundation for trustworthy AI.

Lessons learned:

• Vector-based RAG models work to a degree, but fail to meet expectations for critical applications. The precision paradox applies: as accuracy improves, user dissatisfaction with the remaining errors grows.
• Knowledge Graphs restore trustworthiness by providing deterministic, fact-based reasoning and preserving full content context. They support inferencing (discovering implicit relationships), enable retrieval of related but non-adjacent topics (reducing hallucinations), and scale through automated refreshes. They also handle diverse formats like Markdown, DITA, and AsciiDoc.

At Avalara, they relied on the DITA Map Schema to build a knowledge graph. One needs to be careful as when someone brings content into a vector database, it loses the metadata.

Michael Iantosca proposes a five-step approach: Start with metadata-rich, componentized content Apply taxonomy Build an ontology Convert to RDF Load and refresh in a graph database with inferencing

As Michael highlights, knowledge graphs can replace vector databases for deterministic AI or augment them for hybrid retrieval. Long-term, “structure plus structure” (modular content + knowledge graphs) enables neuro-adaptive documentation and advanced semantic AI.

Takeaway: Trustworthy AI depends less on bigger models and more on structured, contextualized content foundations. Knowledge Graphs offer a path to scalable, accurate, explainable, and maintainable AI systems.