Graph-Augmented Language Models for Strategic Reasoning: A Hybrid Retrieval Framework

Large language models have demonstrated exceptional general language understanding but often struggle with specialized strategic reasoning tasks, such as real-time strategy games. This work introduces a novel retrieval-augmented framework for strategic decision-making in StarCraft II. Our approach constructs a domain-specific knowledge graph (KG) in Neo4j and a semantic vector database of strategic guides (using ChromaDB), providing both symbolic and textual knowledge to the LLM. Given a game-state query, the system retrieves relevant KG triples and strategy-text embeddings via dense similarity search, concatenating them into a unified prompt for the LLM. In experiments, our system significantly outperforms baseline LLMs: it achieves higher strategy accuracy, generates richer action plans, and yields more interpretable decision chains. These results demonstrate that our hybrid KG-and-vector retrieval pipeline enhances LLM reasoning in complex, environments like StarCraft II.