PriHA adds tri-stage RAG for Hong Kong care

This reads less like a flashy assistant launch and more like a serious retrieval architecture paper for regulated healthcare. The real contribution is not the chatbot layer but the attempt to make answers auditable when source material is scattered and domain-specific.

• –The tri-stage design is doing real work: query expansion first, then mixed-source retrieval, then context reorganization before generation.
• –That architecture is a sensible fit for healthcare guidance, where a single vague query often needs multiple sub-questions and provenance matters.
• –The paper’s strongest claim is traceability, not just accuracy, which is the right bar for clinical use cases.
• –The main limitation is obvious: results are paper-level, so the practical question is how well this holds up across real patient phrasing, edge cases, and guideline drift.
• –If it generalizes, the pattern is reusable anywhere local rules and fragmented documents make vanilla RAG brittle.