A groundbreaking achievement in indie game development has been made with the successful integration of a local Llama 3.2 model into the RPG Void Runner, giving birth to the ‘Void Caller AI’ – a dynamic Dungeon Master system. This innovative system leverages a lightweight RAG pipeline to generate quests based on real-time server telemetry, including player deaths, item loots, and boss defeats.

The Void Caller AI interprets server events and seamlessly weaves them into the narrative of the quests it generates, providing players with a unique and immersive experience. By operating locally via Ollama on the backend, the system effectively eliminates cloud API costs and reduces latency to a minimum.

A Python backend serves as the bridge between SQLite telemetry and the Llama 3.2 prompt, utilizing a simplified pipeline to retrieve recent server telemetry and construct a prompt with strict JSON formatting constraints. The Llama 3.2 model then proceeds to generate a dynamic PvE extraction quest based on the context provided.

This pioneering approach enables the game to produce quests that are tailored to the players’ experiences, such as generating a bounty quest to avenge a player’s death. The incorporation of local LLMs in game development has the potential to revolutionize the industry, with this project showcasing the exciting possibilities of AI-powered game design.

Photo by Nino Sanger on Pexels
Photos provided by Pexels