A Layout-Aware Hybrid Retrieval-Augmented Generation Framework for Mitigating Hallucination in Turkish Texts: Evidence from the Turkish Official Gazette

Large language models (LLMs) hallucinate, which limits their use in domains with little tolerance for error, such as law. We present a layout-aware hybrid retrieval-augmented generation (RAG) system for answering questions over 30 issues of the Turkish Official Gazette published in July 2025. The pipeline uses IBM Docling for hierarchical layout analysis and indexes content in PostgreSQL/pgvector with a local multilingual embedding model combined with a lexical index; response synthesis is delegated to claude-sonnet-5 under a citation-constrained prompt. On 33 questions written by the first author, the system answered 17 and declined 16 (coverage 51.5%). Among the answered questions, the hallucination rate was 4.2%, and citation accuracy was 81.8% (n = 11); Hit-Rate@8 was 0.786 on the 14 questions with labeled evidence. A zero-shot baseline gave a verifiable answer to one of 33 questions. In a retrieval ablation, replacing flat PDF text extraction with Docling raised Hit-Rate@8 from 0.21 to 0.79, whereas structural chunking and hybrid retrieval showed no consistent additional gain. A single legal expert rated 30 of 33 responses fully correct. Because the evaluation set is small, covers one month, and was rated by one expert, the results are exploratory and apply to this corpus.