Large language models (LLMs) offer a new paradigm for decision making in real-time strategy games. However, closed-source LLMs rely on costly remote APIs, while open-source LLMs require tens of gigabytes of GPU memory and several hours of inference, creating a severe resource bottleneck. To address this issue, we introduce KADR (Knowledge-Augmented Decision Refinement), which augments a fine-tuned small-parameter Qwen model with a Liquipedia-derived domain knowledge graph and tactical texts. Unlike conventional Retrieval-Augmented Generation (RAG) pipelines that start with broad queries, KADR first lets the LLM produce a preliminary action set, then retrieves knowledge based on the actions, and finally outputs refined decisions. We realize this approach on the platform of StarCraft II. Compared with the baseline method, KADR increases the fine-tuned small model LV5 win rate from 0% to 60%, significantly improving resource and population efficiency, which achieves the competitive performance as good as baseline LLMs, while reducing the per-game inference time from more than 7 h on average to just 1.5 h on a single 8 GB GPU.