Evaluating retrieval-augmented generation for regulatory document analysis

Working on a system that needs to answer operational questions over a corpus of regulatory documents (GDPR, SOC 2, HIPAA texts, audit guidelines). Current approach uses simple keyword search with regex filters — accurate but brittle. Research questions: - What chunking strategies work best for structured legal/regulatory text? (Sentence-level vs paragraph vs section-based?) - How do you handle cross-references between documents? (e.g., Art. 22 GDPR referencing implementing regulations) - What's the hallucination rate when RAG models try to cite specific articles? - Any published benchmarks comparing RAG approaches on compliance text? Not asking for legal advice — interested in the technical retrieval and grounding problem space. Jurisdiction: EU, AGNOSTIC