Virtual Human Knowledge Consistency (VHKC) explores how these issues can be addressed through the combination of GraphRAG (Graph-based Retrieval-Augmented Generation) and Nvidia NeMo Guardrails. Instead of treating all information equally, GraphRAG structures knowledge into different contextual layers, such as world knowledge, local events, and personal memory. Combined with deterministic guardrails, this approach helps virtual characters remain consistent with their role, personality, and the evolving state of the environment.
From Fantasy RPGs to Training Simulations
The project focuses on the development of a working proof of concept within a fantasy RPG setting, where NPCs dynamically react to player behaviour while remaining faithful to the game’s lore and narrative structure.
Beyond entertainment, the research also connects directly to applied simulations and conversational training environments, including projects such as AVATALK. In these contexts, virtual humans must maintain coherent conversations and remember contextual information during longer interactions to remain effective as training partners or educational tools.
Practical Research and Reusable Tools
VHKC aims to translate the research into practical outcomes for both education and industry. The project will result in a reusable code library, technical documentation, and a structured workflow that can support future game productions, applied simulations, and AI-driven virtual human research.
By addressing one of the key limitations of current LLM integrations, the project helps lay the foundation for more believable, reliable, and immersive virtual characters across entertainment, education, and professional training environments.