LLM-VC-Dungeon: Reshaping Classic Text Adventure Games with Large Language Models

LLM-VC-Dungeon is a modern text adventure game based on large language models. It solves the problem of heavy command memory burden in traditional text games through natural language input and an AI Dungeon Master mechanism. The project adopts a front-end and back-end separation architecture, supports local LLM operation (e.g., Ollama), balances privacy and offline availability, and maintains the game's determinism and playability. This article will cover aspects such as background, architecture, mechanisms, and deployment.

Text adventure games are one of the earliest types in the history of video games. From Colossal Cave Adventure in 1976 to the Zork series, they have become classics with rich narratives and imagination. However, traditional games have significant barriers: players need to memorize specific command syntax (e.g., "open chest" to open a treasure chest), and the rigid system is unfriendly to beginners, limiting the audience. With the maturity of LLM technology, the possibility of AI understanding natural language input has emerged, and LLM-VC-Dungeon is the practice of this idea.

LLM-VC-Dungeon is a web-based text adventure game with the core concept of "natural language lowers the threshold, deterministic rules ensure the experience". Front-end and back-end separation architecture: the front-end provides a terminal-style interactive interface, while the back-end is responsible for state management and AI reasoning. The biggest innovation is the introduction of an "AI Dungeon Master" layer, where players can describe actions in daily language (e.g., "carefully inspect the glowing box"), and the AI automatically parses the intent and generates feedback.

Front-end: Developed using the Svelte5 framework + TypeScript. The terminal-style interface evokes nostalgia, and the modern tech stack ensures smooth interaction. Back-end: Based on FastAPI, it undertakes three main responsibilities: intent classification (converting natural language input into structured actions), state management (maintaining the game world state), and narrative generation (calling LLM to generate plot descriptions). LLM Integration: The default recommended solution is Ollama local deployment, with advantages including privacy protection (input is not sent to third parties), cost control (no API fees), offline availability, and flexible model options; it also supports configuration options for Google and OpenAI.

The project balances natural language input and game determinism through a three-layer architecture: 1. Intent Parsing Layer: Maps players' natural language to predefined action types (movement, inspection, etc.); 2. Rule Execution Layer: Checks action feasibility and updates state based on deterministic logic; 3. Narrative Generation Layer: Calls LLM to generate plot descriptions. This architecture not only lowers the input threshold but also ensures the credibility and fairness of the game mechanism, while providing rich narratives.

State Persistence: Implements progress saving and loading in JSON format, supporting resumption after interruption; the save path has security restrictions to prevent path traversal attacks. Deployment Methods: Provides one-click deployment via Docker Compose, including three modes (basic mode: front-end and back-end only; full mode: including Ollama; development mode: local debugging). Development and Expansion: The local development guide is clear (back-end uvicorn hot reload, front-end npm run dev); voice interaction functions may be reintroduced in the future (the architecture reserves expansion space).

Technical Insights: 1. The intent understanding + deterministic execution architecture can be applied to scenarios such as customer service robots and smart homes; 2. Local LLMs can support complex interactive applications without relying on cloud services; 3. The integration of retro interfaces and modern technology provides an example for the refactoring of classic software. Summary: LLM-VC-Dungeon solves the usability problem of text adventure games while retaining their core charm; it is a novel experience for enthusiasts and demonstrates the collaborative mode of LLMs and deterministic systems for developers. Future multimodal/voice technologies may spawn a new interactive paradigm.