The thing about trying to turn TTRPG content into something that can be used in a game is that TTRPG content is a book of art.
Most folks who don’t play TTRPGs don’t realize that the content isn’t a set of tables but rather a piece of art. Everything from the cover, the art, the font, and the language makes or breaks the TTRPG content.
When using it as part of a system for creating campaign content, different parts need to be handled separately.
Want to change a “one-shot” to fit into a campaign? The story itself is the first part you have to change.
Want to make it harder? Then you just need the mechanical details of the monsters.
General-purpose RAG doesn’t work. And shoving a 120 MB PDF to get 10KB is a great way to waste time and make Anthropic’s shareholders and Nvidia’s shareholders wealthier.
So the first step is to transform the PDF into a structured document that can be used to make subset queries.
And because TTRPG is no different from any other discipline, it turns out this problem is why SAP and others have moved away from pure RAG toward creating structured documents that the Agent can interact with.
As a computer scientist, what makes me laugh? The original promise of LLMs was that it could read unstructured data, infer structure and make decisions.
And here we are discovering – yet again – that computers are great at operating on structure, not on systems with no structure.
Nikolas Wirth would be laughing.
Leave a Reply