Cache-Enhanced RAG and Graph-RAG for O-RAN

Generative artificial intelligence (AI), particularly Large Language Models (LLMs), are envisioned to be a core enabler for Sixth Generation (6G) wireless networks, powering use cases such as autonomous troubleshooting, intelligent configuration, and real-time decision support. Open Radio Access Network (O-RAN), with its modular and disaggregated architecture, offers the most suitable platform for embedding these AI capabilities. However, the integration of LLMs into O-RAN is constrained by their high inference latency, lack of domain grounding, and resource-intensive retrieval workflows. To address these limitations, this paper proposes a cache-augmented Retrieval-Augmented Generation (RAG) and Graph-RAG framework specifically designed for O-RAN environments. The framework leverages semantic caching at the User Plane Function (UPF) and centralized model orchestration to support low-latency, high-throughput, and contextually accurate responses. The system is evaluated using the ORAN-Bench 13K benchmark, consisting of 13,952 domain-specific queries derived from official specification documents. Results show that cache-enhanced RAG reduces average latency by 35.8%, increases cache utilization to 76.38%, and achieves a cache hit rate of up to 42.5%, significantly outperforming baseline LLM and Graph-RAG configurations in responsiveness. Graph-RAG achieves the highest factual correctness at 71.2%, followed by RAG at 70.5% and LLM at 64.8%. These results demonstrate that semantic caching not only mitigates the performance bottlenecks of LLM-based reasoning in O-RAN, but also paves the way for scalable, AI-native telecommunication infrastructures aligned with the demands of next-generation networks.