EliteDangerousKB: Practice of Building a Game Knowledge Engine and AI Assistant Based on Local RAG

EliteDangerousKB is an autonomous knowledge engine and local RAG (Retrieval-Augmented Generation) AI assistant built for the game Elite: Dangerous. The project aims to address issues such as low efficiency, poor timeliness in traditional guide information retrieval, and hallucinations of general-purpose LLMs. It ensures data privacy through fully local operation and provides accurate, traceable answers to game-related knowledge queries.

Elite: Dangerous is a deep space simulation game with complex mechanics and continuously updated content. Players face challenges in knowledge acquisition: traditional guides require manual searching which is prone to omissions and outdated content; general-purpose LLMs tend to fabricate incorrect information. EliteDangerousKB addresses these issues through a specialized knowledge base and local RAG technology.

The project adopts a modular architecture with core components including:

• Knowledge Base Module (kb/): Stores structured game knowledge and ensures controllable accuracy;
• Copilot System (copilot/): COVAS voice interaction interface that handles natural language queries;
• Embedding Module (embeddings/): Converts text into vector embeddings to support semantic retrieval;
• Index System (indexes/): Manages vector indexes for efficient retrieval;
• Evaluation Module (eval/): Assesses answer quality to support self-verification capabilities; Additionally, there are subsystems like the Commander Module, Log System, and Real-time Module.

The RAG technology workflow includes:

1. Query Understanding: Analyze user intent and needs;
2. Knowledge Retrieval: Perform vector similarity search for relevant knowledge base fragments;
3. Context Construction: Integrate retrieved content and queries to generate prompts;
4. Answer Generation: LLM generates answers based on context to reduce hallucination risks. The project emphasizes 100% local inference, with advantages including zero latency, privacy protection, offline availability, controllable costs, and configurable models.

The system has an autonomous verification mechanism:

• Answer Consistency Check: Cross-verify multi-source information, assign confidence levels, and provide source traceability;
• Knowledge Base Update: Monitor official logs to automatically extract new content, community crowdsourced verification, and version management;
• AI Committee: May adopt multi-model voting or hierarchical verification to improve answer reliability.

It provides the following value to players:

• Instant Q&A: Get accurate answers about game mechanics, configurations, routes, etc., anytime during gameplay;
• Decision Support: Assist with ship purchases, mission selection, and exploration planning;
• Learning Assistance: New players can learn game knowledge through dialogue;
• Offline Reference: Usable even in network-free environments;
• Customized Experience: Customize content and answer styles based on players' preferences and playstyles.

Technical Insights:

• Specialized knowledge bases are more reliable than general-purpose LLM pre-trained knowledge and can be extended to vertical domains;
• RAG improves answer accuracy and interpretability;
• The development of open-source LLMs makes local deployment a reality, suitable for privacy or network-constrained scenarios;
• Use open-source tools (FAISS, Sentence-BERT, llama.cpp, etc.) to build an ecosystem. Future Directions: Expand to other complex games, technical documentation, education, and customer service fields; add multi-modal support, community contributions, personalized recommendations, voice interaction, and other features.

EliteDangerousKB demonstrates the successful application of RAG technology in the gaming field. Through a specialized knowledge base, RAG architecture, and local deployment, it provides players with a reliable and private AI assistant. The project offers a replicable architecture template, proving how to combine LLMs with domain knowledge to build a trustworthy AI system. It has reference value for solving AI hallucination issues and serves as a research case for developers applying AI to vertical domains.