Serverless AI Game Ecosystem: Practice of Integrated Architecture with Flutter and RAG

The llm-flutter-boilerplate project showcases a high-performance serverless AI game ecosystem, integrating the Flutter engine, Python RAG pipeline, and Gemini model. Its core goal is to solve the hallucination problem in AI game narratives—by using RAG technology to anchor the game knowledge base, it ensures narrative consistency and delivers dynamic, personalized gaming experiences. The architecture adopts a dual-core design, separating the game engine and AI system to balance cross-platform deployment and intelligent narrative capabilities.

The game industry is undergoing a generative AI-driven transformation. Traditional narratives rely on pre-written storylines, but AI intervention enables dynamic and personalized experiences. However, AI-generated plots are prone to hallucinations due to the lack of worldview constraints. This project uses RAG technology to anchor AI generation to the game knowledge base, balancing creativity and narrative consistency.

• Cycull: Flutter game engine layer, responsible for rendering, input processing, and state management. It supports cross-platform deployment on iOS/Android/Web and ensures a smooth 60fps experience.
• Woden: RAG pipeline system built with Python, using Gemini 3 Flash as the reasoning engine. It constrains output through context injection and maintains game world settings, character backgrounds, and plot rules. The separated architecture allows independent iteration and optimization of both cores, focusing on their respective core capabilities.

1. Knowledge Structuring: Uses nested JSON format + metadata annotation to handle complex layers of the game worldview (universe settings, race history, etc.).
2. Context Injection Strategy: Dynamically selects relevant knowledge fragments based on game scenarios and player status, presenting them to the model in a structured way.
3. Consistency Assurance: Maintains a memory bank of "occurred events" to avoid generating content that conflicts with established plots, solving the AI "amnesia" problem.

• Player authentication to protect progress and preferences.
• API access control to limit call frequency and prevent abuse.
• Content filtering to review AI-generated content.
• Data isolation to ensure no interference between player data. The zero-trust mechanism is crucial for public game services—all requests require permission verification.

• Flutter: Near-native performance + unified codebase, suitable for independent teams to iterate quickly.
• Gemini 3 Flash: Lightweight model with fast reasoning speed and low cost, meeting the real-time response needs of games.
• Python: Rich RAG tool ecosystem (vector databases, text processing) supports rapid prototyping.
• Firebase: Serverless architecture eliminates operational burdens, adapting to the needs of small teams.

• Interactive Novels: AI generates dynamic plots consistent with the worldview.
• Educational Games: Generate personalized learning content based on outlines.
• Virtual Worlds: NPCs have long-term memory and consistent behavior.
• Brand Experiences: AI assistants interact strictly in accordance with brand guidelines. The project's boilerplate code provides a starting point for customized development.

The llm-flutter-boilerplate demonstrates a new trend in AI applications: combining the general capabilities of large models with domain knowledge constraints, and solving the hallucination problem by anchoring the knowledge base via RAG. The project provides a complete reference architecture (frontend, backend, AI pipeline, security), offering insights for developers to build AI-driven applications. In the future, multimodal and real-time reasoning technologies will promote the wider application of such architectures.