Can hybrid semantic-graph RAG bridge retrieval and reasoning in Economic Bulletin drafting?

Drafting a comprehensive short-term macroeconomic report requires combining institutional sources, market information, statistical releases and expert knowledge of the macroeconomic landscape. For the purposes of Banca d'Italia's Economic Bulletins, fluency is not sufficient: the drafting process must also ensure factual accuracy, traceability, consistency and reproducibility. Although large language models can accelerate macroeconomic writing, they remain exposed to hallucinations, unsupported claims and weak grounding in verifiable evidence, limiting their use in institutional communication. This paper presents a reproducible retrieval-augmented generation pipeline for Economic Bulletin drafting. The framework combines a semantic vector store with a Neo4j knowledge graph and compares vector, graph and hybrid retrieval modes across chapter-level drafting tasks. The pipeline separates evidence collection, normalization, planning, deterministic drafting, verification, semantic support evaluation, optional open-source refinement and final LaTeX assembly. The results show that vector retrieval is the strongest baseline in most chapters, graph retrieval becomes reliable after schema-aware fallback, and hybrid retrieval is promising but does not automatically dominate under the current selection policy. The contribution is an auditable architecture for controlled, evidence-grounded economic drafting.