Modern telecommunications networks generate huge amounts of operational data. This data needs real-time analysis through Complex Event Processing (CEP) systems. Traditional CEP systems use static rules created by experts. These rules struggle to adapt to changing network conditions. They also create significant maintenance overhead. This thesis investigates using AI to automate CEP pattern creation. It compares two approaches: Retrieval-Augmented Generation (RAG) and Graph-based RAG (GraphRAG) for generating Apache Flink CEP applications. The research creates a systematic taxonomy of CEP patterns. It covers four complexity categories: filtering, aggregation, temporal, and correlation patterns. Two AI-agent architectures were implemented using Claude Sonnet 3.7. Each uses different knowledge retrieval strategies. RAG uses vector-based document retrieval from telecommunications database documentation. GraphRAG uses structured knowledge graphs built in Neo4j. This provides organized metadata and relationship information. A comprehensive evaluation framework was developed. It uses Pass@k metrics combined with multi-dimensional quality assessment. The assessment covers compilation success, runtime execution, and functional correctness. Controlled experiments generated 30 code samples. These included 3 patterns, 2 approaches, and 5 samples each. The samples were executed in a dedicated Kubernetes environment running Apache Flink 1.20.x with real telecommunications data. RAG outperformed GraphRAG overall (86.7% vs. 66.7% Pass@3), mainly because GraphRAG failed on correlation patterns, though it excelled in filtering/aggregation (100% Pass@1). Both performed similarly on temporal patterns, and multisample generation notably improved results (RAG +33.3%). Success varied with pattern complexity: simple filtering was handled well, but correlation patterns remained challenging. RAG’s document-based flexibility proved more resilient, while GraphRAG’s structured design was strong on supported tasks but brittle under complex demands.
Tomas Weldetinsae (Wed,) studied this question.